Skip to content

BrainAPI is a knowledge graph–powered AI memory layer that transforms unstructured data into structured knowledge, enabling intelligent search, recommendations, and contextual memory for AI agents and applications.

Document search architecture and benchmark results

Christian618f2ad43a81
diff --git a/benchmarks/REPORTS.json b/benchmarks/REPORTS.json
index 94a721a..858c569 100644
--- a/benchmarks/REPORTS.json
+++ b/benchmarks/REPORTS.json
@@ -1,6 +1,6 @@
{
"schema_version": 2,
- "updated_at": "2026-08-05T20:07:44.710028+00:00",
+ "updated_at": "2026-08-21T07:53:39.332214+00:00",
"description": "BrainAPI benchmark results. Top published scores across suites. Updated when a suite evaluate/report completes successfully.",
"benchmarks": {
"locomo": {
@@ -803,6 +803,1790 @@
"recorded_at": "2026-08-05T20:07:44.707984+00:00"
}
]
+ },
+ "search": {
+ "name": "Search (hybrid BM25 + dense)",
+ "leaderboard": [
+ {
+ "run_id": "search-20260818T111250Z",
+ "brain": "searchbenchsmoke",
+ "dataset": "search_toy.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "n_queries": 10,
+ "ndcg@10": 1.0,
+ "recall@10": 1.0,
+ "mrr": 1.0,
+ "p50_retrieve_ms": 25.1575,
+ "p95_retrieve_ms": 27.67875,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-20260818T111250Z/report.json",
+ "recorded_at": "2026-08-18T11:13:57.784280+00:00"
+ },
+ {
+ "run_id": "search-toy-cc",
+ "brain": "searchbenchsmoke",
+ "dataset": "search_toy.jsonl",
+ "fusion": "cc",
+ "rerank": "none",
+ "n_queries": 10,
+ "ndcg@10": 1.0,
+ "recall@10": 1.0,
+ "mrr": 1.0,
+ "p50_retrieve_ms": 24.2875,
+ "p95_retrieve_ms": 41.931949999999986,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-toy-cc/report.json",
+ "recorded_at": "2026-08-18T11:15:19.039999+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-allcols",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages",
+ "entities",
+ "communities"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.6815199154995802,
+ "recall@10": 0.4656565656565656,
+ "mrr": 0.8484848484848485,
+ "p50_retrieve_ms": 108.159,
+ "p95_retrieve_ms": 129.32150000000001,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-allcols/report.json",
+ "recorded_at": "2026-08-18T13:02:36.140118+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-passages",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.7579469533596556,
+ "recall@10": 0.5106060606060606,
+ "mrr": 0.8484848484848485,
+ "p50_retrieve_ms": 29.279,
+ "p95_retrieve_ms": 153.97,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-passages/report.json",
+ "recorded_at": "2026-08-18T13:02:47.533040+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-entities",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "entities"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.5323459350114446,
+ "recall@10": 0.4151515151515151,
+ "mrr": 0.6673553719008264,
+ "p50_retrieve_ms": 52.75,
+ "p95_retrieve_ms": 75.479,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-entities/report.json",
+ "recorded_at": "2026-08-18T13:31:39.862006+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-communities",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "communities"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.0,
+ "recall@10": 0.0,
+ "mrr": 0.0,
+ "p50_retrieve_ms": 51.536,
+ "p95_retrieve_ms": 58.432500000000005,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-communities/report.json",
+ "recorded_at": "2026-08-18T13:31:42.972981+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-pe",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages",
+ "entities"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.6815199154995802,
+ "recall@10": 0.4656565656565656,
+ "mrr": 0.8484848484848485,
+ "p50_retrieve_ms": 74.312,
+ "p95_retrieve_ms": 91.006,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-pe/report.json",
+ "recorded_at": "2026-08-18T13:31:46.700938+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-entities-after",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "entities"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.6549492067418874,
+ "recall@10": 0.4818181818181818,
+ "mrr": 0.7568181818181817,
+ "p50_retrieve_ms": 51.071,
+ "p95_retrieve_ms": 74.023,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-entities-after/report.json",
+ "recorded_at": "2026-08-18T13:36:54.694299+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-communities-after",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "communities"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.4776363794553365,
+ "recall@10": 0.35000000000000003,
+ "mrr": 0.8484848484848485,
+ "p50_retrieve_ms": 129.202,
+ "p95_retrieve_ms": 152.889,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-communities-after/report.json",
+ "recorded_at": "2026-08-18T13:39:09.162227+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-passages-smoke",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.7579469533596556,
+ "recall@10": 0.5106060606060606,
+ "mrr": 0.8484848484848485,
+ "p50_retrieve_ms": 27.735,
+ "p95_retrieve_ms": 32.483000000000004,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-passages-smoke/report.json",
+ "recorded_at": "2026-08-18T13:37:03.038349+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-pec-after",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages",
+ "entities",
+ "communities"
+ ],
+ "n_queries": 11,
+ "ndcg@10": 0.6526549564168582,
+ "recall@10": 0.496969696969697,
+ "mrr": 0.816017316017316,
+ "p50_retrieve_ms": 183.946,
+ "p95_retrieve_ms": 216.89249999999998,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-pec-after/report.json",
+ "recorded_at": "2026-08-18T13:39:52.408131+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-passages-ce",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "plugin:cross-encoder",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 11,
+ "ndcg@10": 0.7375861975343849,
+ "ndcg@20": null,
+ "ndcg": null,
+ "recall@10": 0.5106060606060606,
+ "recall@20": 0.8469696969696969,
+ "mrr": 0.8030303030303031,
+ "p50_retrieve_ms": 30.029,
+ "p95_retrieve_ms": 76.32499999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-passages-ce/report.json",
+ "recorded_at": "2026-08-18T14:20:28.184907+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-passages-pool",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": true,
+ "n_queries": 11,
+ "ndcg@10": 0.7884517684883318,
+ "ndcg@20": 0.8114424261982733,
+ "ndcg": 0.8137804024940007,
+ "recall@10": 0.5348484848484848,
+ "recall@20": 0.8469696969696969,
+ "mrr": 0.9090909090909091,
+ "p50_retrieve_ms": 22.721,
+ "p95_retrieve_ms": 75.3825,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-passages-pool/report.json",
+ "recorded_at": "2026-08-18T14:18:44.171034+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-ce-pool",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "none",
+ "rerank": "plugin:cross-encoder",
+ "channels": [
+ "rank-pool-ce"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 11,
+ "ndcg@10": 0.5976277406374534,
+ "ndcg@20": 0.668518640484444,
+ "ndcg": 0.7862048091602294,
+ "recall@10": 0.398637349324982,
+ "recall@20": 0.6598423817505625,
+ "mrr": 0.8333333333333333,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-ce-pool/report.json",
+ "recorded_at": "2026-08-18T14:20:28.034427+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ce-pool",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "plugin:cross-encoder",
+ "channels": [
+ "rank-pool-ce"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 74,
+ "ndcg@10": 0.5492314171926193,
+ "ndcg@20": 0.6311374744057928,
+ "ndcg": 0.7174501167408996,
+ "recall@10": 0.4907400145298616,
+ "recall@20": 0.7577102510384188,
+ "mrr": 0.8306570306570307,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ce-pool/report.json",
+ "recorded_at": "2026-08-18T14:29:17.114211+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ce-pool-ft",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "plugin:cross-encoder",
+ "channels": [
+ "rank-pool-ce"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 74,
+ "ndcg@10": 0.5832632846066277,
+ "ndcg@20": 0.6690889170924552,
+ "ndcg": 0.7451408010463653,
+ "recall@10": 0.4893954738298456,
+ "recall@20": 0.7641337775411228,
+ "mrr": 0.8932519057519057,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ce-pool-ft/report.json",
+ "recorded_at": "2026-08-18T14:32:48.672213+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "ndcg@10": 0.49528253069041633,
+ "ndcg@20": 0.5490094202707876,
+ "ndcg": 0.5490094202707876,
+ "recall@10": 0.3769307112873572,
+ "recall@20": 0.5920965629425753,
+ "mrr": 0.7580598455598455,
+ "p50_retrieve_ms": 70.9225,
+ "p95_retrieve_ms": 94.17495,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages/report.json",
+ "recorded_at": "2026-08-18T15:38:50.867049+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-pool",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "ndcg@10": 0.5458756376131827,
+ "ndcg@20": 0.5785048208487761,
+ "ndcg": 0.6013933046523523,
+ "recall@10": 0.42064799804369324,
+ "recall@20": 0.5920965629425753,
+ "mrr": 0.8254504504504505,
+ "p50_retrieve_ms": 65.447,
+ "p95_retrieve_ms": 91.44999999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-pool/report.json",
+ "recorded_at": "2026-08-18T15:41:37.970033+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-passages-gate",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 11,
+ "ndcg@10": 0.7579469533596556,
+ "ndcg@20": 0.7962544743266965,
+ "ndcg": 0.7962544743266965,
+ "recall@10": 0.5106060606060606,
+ "recall@20": 0.8469696969696969,
+ "mrr": 0.8484848484848485,
+ "p50_retrieve_ms": 25.549,
+ "p95_retrieve_ms": 41.532,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-passages-gate/report.json",
+ "recorded_at": "2026-08-18T16:47:11.911155+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-pec-lists",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages",
+ "entities",
+ "communities"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 11,
+ "ndcg@10": 0.6526549564168582,
+ "ndcg@20": 0.6261389422887412,
+ "ndcg": 0.683063799169122,
+ "recall@10": 0.496969696969697,
+ "recall@20": 0.6530303030303031,
+ "mrr": 0.816017316017316,
+ "p50_retrieve_ms": 160.231,
+ "p95_retrieve_ms": 198.8395,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-pec-lists/report.json",
+ "recorded_at": "2026-08-18T16:47:26.669145+00:00"
+ },
+ {
+ "run_id": "search-esci-slice-entities-neighbors",
+ "brain": "searchbenchesci20",
+ "dataset": "search_esci_slice.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "entities"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 11,
+ "ndcg@10": 0.6549492067418874,
+ "ndcg@20": 0.7173885785786568,
+ "ndcg": 0.7173885785786568,
+ "recall@10": 0.4818181818181818,
+ "recall@20": 0.8217171717171716,
+ "mrr": 0.7568181818181817,
+ "p50_retrieve_ms": 223.289,
+ "p95_retrieve_ms": 271.031,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-slice-entities-neighbors/report.json",
+ "recorded_at": "2026-08-18T16:48:04.631525+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-control",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "ndcg@10": 0.49528253069041633,
+ "ndcg@20": 0.5490094202707876,
+ "ndcg": 0.5490094202707876,
+ "recall@10": 0.3769307112873572,
+ "recall@20": 0.5920965629425753,
+ "mrr": 0.7580598455598455,
+ "p50_retrieve_ms": 61.7115,
+ "p95_retrieve_ms": 91.52104999999995,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-control/report.json",
+ "recorded_at": "2026-08-18T17:13:00.463193+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ce-pool-l12-4class",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "4class-weighted-ce",
+ "channels": [
+ "rank-pool-4class"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 74,
+ "ndcg@10": 0.5968432958924662,
+ "ndcg@20": 0.6782580445181766,
+ "ndcg": 0.7467109659912253,
+ "recall@10": 0.5088965287461584,
+ "recall@20": 0.7673275730894211,
+ "mrr": 0.8351464398761695,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ce-pool-l12-4class/report.json",
+ "recorded_at": "2026-08-18T17:44:17.593025+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ce-pool-4class",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "4class-weighted-ce",
+ "channels": [
+ "rank-pool-4class"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 74,
+ "ndcg@10": 0.5826375707951627,
+ "ndcg@20": 0.6696127031036031,
+ "ndcg": 0.7419804620261266,
+ "recall@10": 0.49798997460620914,
+ "recall@20": 0.7668403956654921,
+ "mrr": 0.8317567567567568,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ce-pool-4class/report.json",
+ "recorded_at": "2026-08-18T17:54:14.073711+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ce-pool-4class-nowt",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "4class-weighted-ce",
+ "channels": [
+ "rank-pool-4class"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 74,
+ "ndcg@10": 0.6096378704103984,
+ "ndcg@20": 0.6881968860231006,
+ "ndcg": 0.7594213054854562,
+ "recall@10": 0.5249760000665886,
+ "recall@20": 0.7631438693842529,
+ "mrr": 0.8530036338859869,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ce-pool-4class-nowt/report.json",
+ "recorded_at": "2026-08-18T18:20:20.750483+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ce-pool-4class-e2",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "4class-weighted-ce",
+ "channels": [
+ "rank-pool-4class"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 74,
+ "ndcg@10": 0.6169746541917066,
+ "ndcg@20": 0.6945278030601721,
+ "ndcg": 0.7652624576380807,
+ "recall@10": 0.5214445444747297,
+ "recall@20": 0.7719306369596903,
+ "mrr": 0.8525418275418275,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ce-pool-4class-e2/report.json",
+ "recorded_at": "2026-08-18T18:48:00.939749+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4998984443226263,
+ "ndcg@20": 0.5522842574567189,
+ "ndcg@50": 0.6378451547901237,
+ "ndcg": 0.6378451547901237,
+ "recall@10": 0.378732513089159,
+ "recall@20": 0.589865760711773,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7645849420849421,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 60.091,
+ "p95_retrieve_ms": 88.02255,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50/report.json",
+ "recorded_at": "2026-08-18T20:31:55.998500+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k100",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 100,
+ "ndcg@10": 0.4998055248060819,
+ "ndcg@20": 0.5523315287868084,
+ "ndcg@50": 0.6375525242924566,
+ "ndcg": 0.6564185982755066,
+ "recall@10": 0.378732513089159,
+ "recall@20": 0.5913672622132746,
+ "recall@50": 0.830951367545064,
+ "mrr": 0.76497727997728,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 57.553,
+ "p95_retrieve_ms": 81.81569999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k100/report.json",
+ "recorded_at": "2026-08-18T20:32:27.553496+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50-pool",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5668024300832123,
+ "ndcg@20": 0.624102711461205,
+ "ndcg@50": 0.6826188564511254,
+ "ndcg": 0.6947261855882534,
+ "recall@10": 0.4507852781809733,
+ "recall@20": 0.6686543312372134,
+ "recall@50": 0.8324528690465655,
+ "mrr": 0.8271396396396397,
+ "pool_coverage": 0.778222315718655,
+ "p50_retrieve_ms": 60.1315,
+ "p95_retrieve_ms": 86.49954999999997,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50-pool/report.json",
+ "recorded_at": "2026-08-18T20:33:02.203247+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k100-pool",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 100,
+ "ndcg@10": 0.5737527941225129,
+ "ndcg@20": 0.6352462971440838,
+ "ndcg@50": 0.7092403328584962,
+ "ndcg": 0.7183777959320791,
+ "recall@10": 0.46580029319598837,
+ "recall@20": 0.6964864396614271,
+ "recall@50": 0.9062083174986402,
+ "mrr": 0.836599099099099,
+ "pool_coverage": 0.8462618325454071,
+ "p50_retrieve_ms": 58.5245,
+ "p95_retrieve_ms": 82.5653,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k100-pool/report.json",
+ "recorded_at": "2026-08-18T20:33:39.620042+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50-ce",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "harness:cross-encoder",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4478964444706324,
+ "ndcg@20": 0.5146313751051845,
+ "ndcg@50": 0.6060742524325241,
+ "ndcg": 0.6060742524325241,
+ "recall@10": 0.331273452580599,
+ "recall@20": 0.5538599992903688,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7197929141477528,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 60.091,
+ "p95_retrieve_ms": 88.02255,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50-ce/report.json",
+ "recorded_at": "2026-08-18T20:34:39.854467+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50-ce-l12",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "harness:cross-encoder",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4674585131694284,
+ "ndcg@20": 0.5084668214514952,
+ "ndcg@50": 0.6140738833603318,
+ "ndcg": 0.6140738833603318,
+ "recall@10": 0.3412172585548375,
+ "recall@20": 0.5191872363321448,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7063349063349063,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 60.091,
+ "p95_retrieve_ms": 88.02255,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50-ce-l12/report.json",
+ "recorded_at": "2026-08-18T20:35:44.462471+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50-qrewrite",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74_qrewrite.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4998984443226263,
+ "ndcg@20": 0.5522842574567189,
+ "ndcg@50": 0.6378451547901237,
+ "ndcg": 0.6378451547901237,
+ "recall@10": 0.378732513089159,
+ "recall@20": 0.589865760711773,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7645849420849421,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 59.432500000000005,
+ "p95_retrieve_ms": 82.09889999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50-qrewrite/report.json",
+ "recorded_at": "2026-08-18T21:18:51.308816+00:00"
+ },
+ {
+ "run_id": "search-esci-74-dense-ance-k50",
+ "brain": "harness-local",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "none",
+ "channels": [
+ "harness-dense"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.3960504272587966,
+ "ndcg@20": 0.445560860839057,
+ "ndcg@50": 0.5370854731938006,
+ "ndcg": 0.5370854731938006,
+ "recall@10": 0.3015013697338196,
+ "recall@20": 0.4872087291019431,
+ "recall@50": 0.7503156876126188,
+ "mrr": 0.6471417846417846,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 74.9370945944778,
+ "p95_retrieve_ms": 74.9370945944778,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-dense-ance-k50/report.json",
+ "recorded_at": "2026-08-18T21:22:24.054560+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50-ce-retrieved",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "harness:cross-encoder",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4648895994453083,
+ "ndcg@20": 0.5245865140860362,
+ "ndcg@50": 0.6126753191137789,
+ "ndcg": 0.6126753191137789,
+ "recall@10": 0.35140463046557896,
+ "recall@20": 0.5410788442403298,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7120610588001892,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 60.091,
+ "p95_retrieve_ms": 88.02255,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50-ce-retrieved/report.json",
+ "recorded_at": "2026-08-18T22:38:31.141079+00:00"
+ },
+ {
+ "run_id": "search-esci-74-colbert-k50",
+ "brain": "harness-local-colbert",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "none",
+ "channels": [
+ "harness-colbert"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.43412701768229023,
+ "ndcg@20": 0.4815601734708754,
+ "ndcg@50": 0.5523204205368764,
+ "ndcg": 0.5523204205368764,
+ "recall@10": 0.3110168068124024,
+ "recall@20": 0.5023603484326227,
+ "recall@50": 0.7138820979266567,
+ "mrr": 0.6689285195333583,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 10063.364791662614,
+ "p95_retrieve_ms": 10063.364791662614,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-colbert-k50/report.json",
+ "recorded_at": "2026-08-18T22:51:16.979167+00:00"
+ },
+ {
+ "run_id": "search-esci-74-bge-base-k50",
+ "brain": "harness-local",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "none",
+ "channels": [
+ "harness-dense"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.44136708015333975,
+ "ndcg@20": 0.5041842920107446,
+ "ndcg@50": 0.5892299329783445,
+ "ndcg": 0.5892299329783445,
+ "recall@10": 0.32179328339461255,
+ "recall@20": 0.5251414009467592,
+ "recall@50": 0.7989247943561789,
+ "mrr": 0.7471303646913403,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 408.0482843509762,
+ "p95_retrieve_ms": 408.04828435097625,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-bge-base-k50/report.json",
+ "recorded_at": "2026-08-18T22:52:53.742680+00:00"
+ },
+ {
+ "run_id": "search-esci-74-union-bge-k50",
+ "brain": "harness-local-union",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "harness-union"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.47918374395281466,
+ "ndcg@20": 0.5381296158524957,
+ "ndcg@50": 0.6141255447421892,
+ "ndcg": 0.6141255447421892,
+ "recall@10": 0.36314016383757297,
+ "recall@20": 0.5774058852814317,
+ "recall@50": 0.8271070974286027,
+ "mrr": 0.7653163449499656,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-union-bge-k50/report.json",
+ "recorded_at": "2026-08-19T08:10:06.928384+00:00"
+ },
+ {
+ "run_id": "search-esci-74-exhaustive-ce",
+ "brain": "harness-local-exhaustive",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "4class-weighted-ce",
+ "channels": [
+ "exhaustive-4class"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4162185585932262,
+ "ndcg@20": 0.4392534244105597,
+ "ndcg@50": 0.5042093175793891,
+ "ndcg": 0.5042093175793891,
+ "recall@10": 0.26585896457775815,
+ "recall@20": 0.4060812399456508,
+ "recall@50": 0.6029270278490679,
+ "mrr": 0.6578079129936715,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 94253.11211881155,
+ "p95_retrieve_ms": 94253.11211881155,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-exhaustive-ce/report.json",
+ "recorded_at": "2026-08-19T10:05:59.463503+00:00"
+ },
+ {
+ "run_id": "search-esci-74-catalog-ce-k50",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "plugin:cross-encoder",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4674585131694284,
+ "ndcg@20": 0.5084668214514952,
+ "ndcg@50": 0.6140738833603318,
+ "ndcg": 0.6140738833603318,
+ "recall@10": 0.3412172585548375,
+ "recall@20": 0.5191872363321448,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7063349063349063,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 59.308499999999995,
+ "p95_retrieve_ms": 86.88884999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-catalog-ce-k50/report.json",
+ "recorded_at": "2026-08-19T10:17:45.568452+00:00"
+ },
+ {
+ "run_id": "search-esci-74-cascade-tail-k50",
+ "brain": "harness-local-cascade",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "none",
+ "channels": [
+ "harness-cascade"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4998984443226263,
+ "ndcg@20": 0.6194622913116794,
+ "ndcg@50": 0.6686100079902446,
+ "ndcg": 0.6686100079902446,
+ "recall@10": 0.378732513089159,
+ "recall@20": 0.7504072036031975,
+ "recall@50": 0.8885964311815054,
+ "mrr": 0.7679931554931555,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-cascade-tail-k50/report.json",
+ "recorded_at": "2026-08-19T10:32:17.738702+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ce-pool-deberta-base",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "4class-weighted-ce",
+ "channels": [
+ "rank-pool-4class"
+ ],
+ "rank_pool": true,
+ "rank_pool_ce": true,
+ "n_queries": 74,
+ "k": 40,
+ "ndcg@10": 0.6492728519674476,
+ "ndcg@20": 0.7101117922903459,
+ "ndcg@50": null,
+ "ndcg": 0.7772540710194906,
+ "recall@10": 0.5183145850778,
+ "recall@20": 0.7604145345598889,
+ "recall@50": null,
+ "mrr": 0.8646712599298806,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ce-pool-deberta-base/report.json",
+ "recorded_at": "2026-08-19T14:00:25.594211+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50-ce-deberta",
+ "brain": "searchbenchsmoke",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "harness:cross-encoder",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5101053946025641,
+ "ndcg@20": 0.5568741362838021,
+ "ndcg@50": 0.6365973490191127,
+ "ndcg": 0.6365973490191127,
+ "recall@10": 0.36281504203381093,
+ "recall@20": 0.5467544580912761,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7457139784725992,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 60.091,
+ "p95_retrieve_ms": 88.02255,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50-ce-deberta/report.json",
+ "recorded_at": "2026-08-19T14:09:19.540811+00:00"
+ },
+ {
+ "run_id": "search-esci-es-passages-k50",
+ "brain": "searchbenchescies",
+ "dataset": "search_esci_es.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 62,
+ "k": 50,
+ "ndcg@10": 0.5774393479403104,
+ "ndcg@20": 0.6303553323648456,
+ "ndcg@50": 0.7327853549091549,
+ "ndcg": 0.7327853549091549,
+ "recall@10": 0.35328224222955834,
+ "recall@20": 0.5758587932370467,
+ "recall@50": 0.9141280526823616,
+ "mrr": 0.8704493087557603,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 84.13,
+ "p95_retrieve_ms": 96.16295000000001,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-es-passages-k50/report.json",
+ "recorded_at": "2026-08-19T16:34:42.965337+00:00"
+ },
+ {
+ "run_id": "search-italian-smoke",
+ "brain": "searchbenchitsmoke",
+ "dataset": "search_italian_smoke.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 3,
+ "k": 20,
+ "ndcg@10": 1.0,
+ "ndcg@20": 1.0,
+ "ndcg@50": null,
+ "ndcg": 1.0,
+ "recall@10": 1.0,
+ "recall@20": 1.0,
+ "recall@50": null,
+ "mrr": 1.0,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 25.33,
+ "p95_retrieve_ms": 29.3206,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-italian-smoke/report.json",
+ "recorded_at": "2026-08-19T16:37:37.730780+00:00"
+ },
+ {
+ "run_id": "search-italian-smoke-inflect",
+ "brain": "searchbenchitsmoke",
+ "dataset": "search_italian_smoke_inflect.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 3,
+ "k": 20,
+ "ndcg@10": 1.0,
+ "ndcg@20": 1.0,
+ "ndcg@50": null,
+ "ndcg": 1.0,
+ "recall@10": 1.0,
+ "recall@20": 1.0,
+ "recall@50": null,
+ "mrr": 1.0,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 21.494,
+ "p95_retrieve_ms": 23.211199999999998,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-italian-smoke-inflect/report.json",
+ "recorded_at": "2026-08-19T16:37:46.108861+00:00"
+ },
+ {
+ "run_id": "search-italian-smoke-inflect-fts",
+ "brain": "searchbenchitsmoke",
+ "dataset": "search_italian_smoke_inflect.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 3,
+ "k": 20,
+ "ndcg@10": 1.0,
+ "ndcg@20": 1.0,
+ "ndcg@50": null,
+ "ndcg": 1.0,
+ "recall@10": 1.0,
+ "recall@20": 1.0,
+ "recall@50": null,
+ "mrr": 1.0,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 33.56,
+ "p95_retrieve_ms": 36.8936,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-italian-smoke-inflect-fts/report.json",
+ "recorded_at": "2026-08-19T16:44:05.699330+00:00"
+ },
+ {
+ "run_id": "search-italian-minilm-pipeline",
+ "brain": "searchbenchitmmini",
+ "dataset": "search_italian_smoke.jsonl",
+ "fusion": "none",
+ "rerank": "none",
+ "channels": [
+ "harness-dense"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 3,
+ "k": 10,
+ "ndcg@10": 1.0,
+ "ndcg@20": 1.0,
+ "ndcg@50": null,
+ "ndcg": 1.0,
+ "recall@10": 1.0,
+ "recall@20": null,
+ "recall@50": null,
+ "mrr": 1.0,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 188.19916666689096,
+ "p95_retrieve_ms": 188.19916666689096,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-italian-minilm-pipeline/report.json",
+ "recorded_at": "2026-08-19T19:27:33.792776+00:00"
+ },
+ {
+ "run_id": "search-esci-es-spell-k50",
+ "brain": "searchbenchescies",
+ "dataset": "search_esci_es_spell.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 62,
+ "k": 50,
+ "ndcg@10": 0.5950327313726855,
+ "ndcg@20": 0.6464666665333806,
+ "ndcg@50": 0.7424579237775919,
+ "ndcg": 0.7424579237775919,
+ "recall@10": 0.36220476611566177,
+ "recall@20": 0.5832920760052671,
+ "recall@50": 0.9143064382925891,
+ "mrr": 0.8865783410138248,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 92.655,
+ "p95_retrieve_ms": 107.2923,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-es-spell-k50/report.json",
+ "recorded_at": "2026-08-19T19:28:14.068481+00:00"
+ },
+ {
+ "run_id": "search-italian-smoke-extras-locale",
+ "brain": "searchbenchitsmoke",
+ "dataset": "search_italian_smoke.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 3,
+ "k": 10,
+ "ndcg@10": 1.0,
+ "ndcg@20": 1.0,
+ "ndcg@50": null,
+ "ndcg": 1.0,
+ "recall@10": 1.0,
+ "recall@20": null,
+ "recall@50": null,
+ "mrr": 1.0,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 17.126,
+ "p95_retrieve_ms": 23.894000000000002,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-italian-smoke-extras-locale/report.json",
+ "recorded_at": "2026-08-19T19:30:47.530480+00:00"
+ },
+ {
+ "run_id": "search-italian-smoke-extras-nope",
+ "brain": "searchbenchitsmoke",
+ "dataset": "search_italian_smoke.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 3,
+ "k": 10,
+ "ndcg@10": 0.0,
+ "ndcg@20": 0.0,
+ "ndcg@50": null,
+ "ndcg": 0.0,
+ "recall@10": 0.0,
+ "recall@20": null,
+ "recall@50": null,
+ "mrr": 0.0,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 22.721,
+ "p95_retrieve_ms": 24.2753,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-italian-smoke-extras-nope/report.json",
+ "recorded_at": "2026-08-19T19:29:31.676418+00:00"
+ },
+ {
+ "run_id": "search-esci-74-passages-k50-extras-sanity",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4998984443226263,
+ "ndcg@20": 0.5522842574567189,
+ "ndcg@50": 0.6378451547901237,
+ "ndcg": 0.6378451547901237,
+ "recall@10": 0.378732513089159,
+ "recall@20": 0.589865760711773,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7645849420849421,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 63.1325,
+ "p95_retrieve_ms": 88.86579999999998,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-passages-k50-extras-sanity/report.json",
+ "recorded_at": "2026-08-19T19:35:44.545077+00:00"
+ },
+ {
+ "run_id": "search-italian-smoke-inflect-fts-or",
+ "brain": "searchbenchitsmoke",
+ "dataset": "search_italian_smoke_inflect.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 3,
+ "k": 20,
+ "ndcg@10": 1.0,
+ "ndcg@20": 1.0,
+ "ndcg@50": null,
+ "ndcg": 1.0,
+ "recall@10": 1.0,
+ "recall@20": 1.0,
+ "recall@50": null,
+ "mrr": 1.0,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 26.364,
+ "p95_retrieve_ms": 42.744899999999994,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-italian-smoke-inflect-fts-or/report.json",
+ "recorded_at": "2026-08-19T20:16:24.241357+00:00"
+ },
+ {
+ "run_id": "search-esci-74-cc-a03",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "cc",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.48296913346675313,
+ "ndcg@20": 0.5353434819101804,
+ "ndcg@50": 0.6235227747628354,
+ "ndcg": 0.6235227747628354,
+ "recall@10": 0.37594464118364007,
+ "recall@20": 0.5859902261910772,
+ "recall@50": 0.8324528690465655,
+ "mrr": 0.7492780202562811,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 62.8885,
+ "p95_retrieve_ms": 89.98204999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-cc-a03/report.json",
+ "recorded_at": "2026-08-19T20:18:18.361436+00:00"
+ },
+ {
+ "run_id": "search-esci-74-cc-a05",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "cc",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4861505982058926,
+ "ndcg@20": 0.5412686017716648,
+ "ndcg@50": 0.62608241624179,
+ "ndcg": 0.62608241624179,
+ "recall@10": 0.37673766109430706,
+ "recall@20": 0.5919104321112831,
+ "recall@50": 0.8324528690465655,
+ "mrr": 0.7604021879021879,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 61.432,
+ "p95_retrieve_ms": 92.56864999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-cc-a05/report.json",
+ "recorded_at": "2026-08-19T20:18:41.196658+00:00"
+ },
+ {
+ "run_id": "search-esci-74-cc-a07",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "cc",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4925393217389386,
+ "ndcg@20": 0.5491258060245199,
+ "ndcg@50": 0.6321844176410597,
+ "ndcg": 0.6321844176410597,
+ "recall@10": 0.3780890124456584,
+ "recall@20": 0.5964401922862047,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7630083655083656,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 60.5155,
+ "p95_retrieve_ms": 89.04719999999998,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-cc-a07/report.json",
+ "recorded_at": "2026-08-19T20:19:05.271923+00:00"
+ },
+ {
+ "run_id": "search-esci-74-literal-fill-k50",
+ "brain": "searchbenchesci74",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.4998984443226263,
+ "ndcg@20": 0.45027838324634667,
+ "ndcg@50": 0.5436008239474691,
+ "ndcg": 0.5436008239474691,
+ "recall@10": 0.378732513089159,
+ "recall@20": 0.3929861145019264,
+ "recall@50": 0.6549969089908488,
+ "mrr": 0.7639237451737452,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 49.019999999999996,
+ "p95_retrieve_ms": 74.27314999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-literal-fill-k50/report.json",
+ "recorded_at": "2026-08-19T20:20:15.131585+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ltr-head-k50",
+ "brain": "harness-local-ltr",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "ltr-pairwise",
+ "channels": [
+ "harness-ltr-head"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5146015046757452,
+ "ndcg@20": 0.5567887112159354,
+ "ndcg@50": 0.6406685515140392,
+ "ndcg": 0.6406685515140392,
+ "recall@10": 0.384165323643092,
+ "recall@20": 0.5830174861607277,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7525579150579151,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ltr-head-k50/report.json",
+ "recorded_at": "2026-08-19T21:09:47.411344+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ltr-cefeat-k50",
+ "brain": "harness-local-ltr",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "ltr-pairwise",
+ "channels": [
+ "harness-ltr-head"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5242780297400812,
+ "ndcg@20": 0.5655195290162985,
+ "ndcg@50": 0.6456032400547804,
+ "ndcg": 0.6456032400547804,
+ "recall@10": 0.38343544674019964,
+ "recall@20": 0.5687827599719494,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7792567567567568,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ltr-cefeat-k50/report.json",
+ "recorded_at": "2026-08-20T07:14:39.761450+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ltr-deberta-k50",
+ "brain": "harness-local-ltr",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "ltr-pairwise",
+ "channels": [
+ "harness-ltr-head"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5420964444237363,
+ "ndcg@20": 0.580013195834899,
+ "ndcg@50": 0.6509103805030941,
+ "ndcg": 0.6509103805030941,
+ "recall@10": 0.3872531182148088,
+ "recall@20": 0.5863055009035371,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7744787644787645,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ltr-deberta-k50/report.json",
+ "recorded_at": "2026-08-20T07:43:07.017767+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ltr-deberta-lgbm-k50",
+ "brain": "harness-local-ltr",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "ltr-lightgbm",
+ "channels": [
+ "harness-ltr-head"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5159432058815566,
+ "ndcg@20": 0.5658120484311862,
+ "ndcg@50": 0.6415336394479464,
+ "ndcg": 0.6415336394479464,
+ "recall@10": 0.37878019077485103,
+ "recall@20": 0.5918308018978374,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7098312215959275,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ltr-deberta-lgbm-k50/report.json",
+ "recorded_at": "2026-08-20T07:48:40.794647+00:00"
+ },
+ {
+ "run_id": "search-esci-ltr200-passages-k50",
+ "brain": "searchbenchesciltr2",
+ "dataset": "search_esci_ltr200.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 170,
+ "k": 50,
+ "ndcg@10": 0.38688497520690734,
+ "ndcg@20": 0.4428785087667844,
+ "ndcg@50": 0.5191838961643084,
+ "ndcg": 0.5191838961643084,
+ "recall@10": 0.3668275379939918,
+ "recall@20": 0.5506847180507305,
+ "recall@50": 0.7435914662537475,
+ "mrr": 0.6599425680048517,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 76.07849999999999,
+ "p95_retrieve_ms": 152.37659999999983,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-ltr200-passages-k50/report.json",
+ "recorded_at": "2026-08-20T10:45:30.571169+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ltr-deberta-train200",
+ "brain": "harness-local-ltr",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "ltr-pairwise",
+ "channels": [
+ "harness-ltr-head"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5436096369070966,
+ "ndcg@20": 0.5825504636416924,
+ "ndcg@50": 0.6525856561608427,
+ "ndcg": 0.6525856561608427,
+ "recall@10": 0.3911684808472767,
+ "recall@20": 0.5850220289937974,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7621138996138996,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ltr-deberta-train200/report.json",
+ "recorded_at": "2026-08-20T10:57:34.082897+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ltr-deberta-lgbm-train200",
+ "brain": "harness-local-ltr",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "ltr-lightgbm",
+ "channels": [
+ "harness-ltr-head"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5333470812074246,
+ "ndcg@20": 0.5739148370361776,
+ "ndcg@50": 0.6521013344978932,
+ "ndcg": 0.6521013344978932,
+ "recall@10": 0.38938710235182716,
+ "recall@20": 0.5909398273461974,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7517797970384177,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ltr-deberta-lgbm-train200/report.json",
+ "recorded_at": "2026-08-20T11:33:18.933082+00:00"
+ },
+ {
+ "run_id": "search-esci-74-ltr-deberta-hybrid170",
+ "brain": "harness-local-ltr",
+ "dataset": "search_esci_74.jsonl",
+ "fusion": "none",
+ "rerank": "ltr-pairwise",
+ "channels": [
+ "harness-ltr-head"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 74,
+ "k": 50,
+ "ndcg@10": 0.5304449022081842,
+ "ndcg@20": 0.5753013243082012,
+ "ndcg@50": 0.6452218148321592,
+ "ndcg": 0.6452218148321592,
+ "recall@10": 0.38518108116293304,
+ "recall@20": 0.5882930387092911,
+ "recall@50": 0.833954370548067,
+ "mrr": 0.7646718146718147,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 0.0,
+ "p95_retrieve_ms": 0.0,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-esci-74-ltr-deberta-hybrid170/report.json",
+ "recorded_at": "2026-08-20T11:57:34.067863+00:00"
+ },
+ {
+ "run_id": "search-wands-passages-k50",
+ "brain": "searchbenchwands",
+ "dataset": "search_wands.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 66,
+ "k": 50,
+ "ndcg@10": 0.823176135584089,
+ "ndcg@20": 0.8120944688355884,
+ "ndcg@50": 0.8223686140560469,
+ "ndcg": 0.8223686140560469,
+ "recall@10": 0.26880378393536286,
+ "recall@20": 0.46518737110842373,
+ "recall@50": 0.8365552576078893,
+ "mrr": 0.9245931537598204,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 86.595,
+ "p95_retrieve_ms": 96.93674999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-wands-passages-k50/report.json",
+ "recorded_at": "2026-08-20T13:17:40.353988+00:00"
+ },
+ {
+ "run_id": "search-wandsgraph-communities-k50",
+ "brain": "searchbenchwandsgraph",
+ "dataset": "search_wands.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "communities"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 66,
+ "k": 50,
+ "ndcg@10": 0.4094084735740818,
+ "ndcg@20": 0.36455789387786824,
+ "ndcg@50": 0.325224291157982,
+ "ndcg": 0.4570937299953597,
+ "recall@10": 0.14824773180036338,
+ "recall@20": 0.20624528979792137,
+ "recall@50": 0.28701452056715215,
+ "mrr": 0.6737914862914863,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 5609.237499999999,
+ "p95_retrieve_ms": 11856.0285,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-wandsgraph-communities-k50/report.json",
+ "recorded_at": "2026-08-21T03:12:00.249304+00:00",
+ "protocol": "architecture demo: isolated catalog graph on searchbenchwandsgraph; not a quality-default claim vs frozen WANDS passages 0.823 or ESCI 0.500",
+ "claim": "architecture-demo"
+ },
+ {
+ "run_id": "search-wandsgraph-passages-k50",
+ "brain": "searchbenchwandsgraph",
+ "dataset": "search_wands.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 66,
+ "k": 50,
+ "ndcg@10": 0.8230998576010691,
+ "ndcg@20": 0.812056754200934,
+ "ndcg@50": 0.8223486981324697,
+ "ndcg": 0.8223486981324697,
+ "recall@10": 0.26880378393536286,
+ "recall@20": 0.46518737110842373,
+ "recall@50": 0.8365552576078893,
+ "mrr": 0.9245931537598204,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 98.604,
+ "p95_retrieve_ms": 115.12925,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-wandsgraph-passages-k50/report.json",
+ "recorded_at": "2026-08-21T03:12:36.455586+00:00",
+ "protocol": "architecture demo: isolated catalog graph on searchbenchwandsgraph; not a quality-default claim vs frozen WANDS passages 0.823 or ESCI 0.500",
+ "claim": "architecture-demo"
+ },
+ {
+ "run_id": "search-wandsgraph-passages-nodeid-k50",
+ "brain": "searchbenchwandsgraph",
+ "dataset": "search_wands.jsonl",
+ "fusion": "rrf",
+ "rerank": "none",
+ "channels": [
+ "passages"
+ ],
+ "rank_pool": false,
+ "rank_pool_ce": false,
+ "n_queries": 66,
+ "k": 50,
+ "ndcg@10": 0.8230998576010691,
+ "ndcg@20": 0.812056754200934,
+ "ndcg@50": 0.8223486981324697,
+ "ndcg": 0.8223486981324697,
+ "recall@10": 0.26880378393536286,
+ "recall@20": 0.46518737110842373,
+ "recall@50": 0.8365552576078893,
+ "mrr": 0.9245931537598204,
+ "pool_coverage": null,
+ "p50_retrieve_ms": 84.0175,
+ "p95_retrieve_ms": 105.24199999999999,
+ "git_sha": "f42c4a8eaadb3c0cc79f78c34b0e1a70534bee10",
+ "report_path": "runs/search-wandsgraph-passages-nodeid-k50/report.json",
+ "recorded_at": "2026-08-21T07:53:39.311033+00:00",
+ "protocol": "architecture demo: isolated catalog graph on searchbenchwandsgraph; not a quality-default claim vs frozen WANDS passages 0.823 or ESCI 0.500",
+ "claim": "architecture-demo"
+ }
+ ]
}
}
}
diff --git a/docs/decisions/007-three-product-surfaces-one-kb.md b/docs/decisions/007-three-product-surfaces-one-kb.md
new file mode 100644
index 0000000..ff4b88f
--- /dev/null
+++ b/docs/decisions/007-three-product-surfaces-one-kb.md
@@ -0,0 +1,128 @@
+# ADR-007: Three Product Surfaces on One Knowledge Base
+
+## Status
+
+Accepted — 2026-08-18. Does not supersede ADR-005 or ADR-006. Those remain the memory-path contract. This ADR adds **search** as a first-class product beside memory and recommendations, and states how they share the same brain without sharing request semantics.
+
+Extended by [ADR-008](008-query-gated-search-personalization.md) (Accepted 2026-08-21): optional search `target` reranks retrieved hits. The omitted-`target` path and this ADR’s SLO are unchanged.
+
+Locked §15 resolutions:
+
+1. Routes stay under `/retrieve` as `GET|POST /retrieve/search` (not a new top-level prefix). Existing `GET /retrieve/` is unchanged.
+2. Search is env-gated (`SEARCH_ENABLED`, also in TUI init/config). v1 lexical is Okapi BM25 over Postgres `tsvector` + GIN on the stock `pgvector/pgvector:pg16` image — not `ILIKE`, not `ts_rank_cd`, not `rum` / ParadeDB `pg_search`.
+3. Dense quality at 3072-d: keep float32 storage; build HNSW on a `halfvec` expression (legal to 4000-d); exact float32 cosine rerank of the overfetch window. Do not truncate embedding dimensions.
+4. `plugins/chatbot-memory` is out of scope for v1.
+5. Search SLO: p50 **< 200 ms excluding embed RTT**. `profile_stages` splits `embed.query` from retrieve so the budget is measurable.
+6. Operators switch dense / BM25 / both via `SEARCH_USE_DENSE` / `SEARCH_USE_BM25`; default is both fused with `SEARCH_FUSION=rrf`. `/retrieve/context` stays ILIKE unless `SEARCH_ENABLED=true` and `CONTEXT_PASSAGE_MODE` is `hybrid` / `bm25` / `dense`.
+
+## Date
+
+2026-08-18
+
+## Context
+
+`docs/research/00-scope-and-constraints.md` already names the product as a memory layer that is also “a substrate for agents, recommenders, and search engines.” In code those are not three products yet:
+
+| Intended product | Shipped surface | What the caller actually gets |
+| --- | --- | --- |
+| Agent / conversational memory | `POST /retrieve/context` | Prompt-shaped blob: passages + event facts + topics. Budget **p50 < 1000 ms**, no LLM (ADR-006). Evaluated on LoCoMo / LongMemEval / BEAM. |
+| Recommendations | `GET|POST /retrieve/recommend` | Ranked **nodes** (items) from graph walks, synergies, optional attribute prefs. Evaluated on `demorecsys` (`16-recsys-eval-protocol.md`). Heavy CF is a plugin (`plugins/recsys-gnn`). |
+| Search | **Missing** | `GET /retrieve/` is entity-biased graph+data lookup. Passage “hybrid” in context is dense ANN ∪ **`ILIKE '%query%'`**, fused with RRF — not BM25, not a hit list API, not a search SLO. |
+
+Default pgvector installs ship **3072-d** embeddings and skip HNSW (`vectors.py` returns no DDL above 2000 dims). Context p50 after dossier deletion is **~1.3–1.4 s**, residual dominated by **serial embedding HTTP**, not ANN. Passage evidence-session recall on LoCoMo conv-26 is already **~97%**; graph vs passages-only does not move judge accuracy above the 5–7 point noise floor.
+
+If search is implemented by stuffing more ranking into `/retrieve/context`, two failures follow: (1) memory benchmarks and the sub-second budget become the search SLO, which they cannot be; (2) search users receive a concatenated context blob instead of scored hits with filters, facets, and stable document IDs.
+
+The literature that actually transfers is first-stage IR (BM25, dense dual-encoders, hybrid fusion) and BEIR’s finding that **lexical and dense are complementary off-distribution** — not MS MARCO leaderboard recipes that assume a 200 ms p99 web index and a cross-encoder on every query. Analysis and task breakdown: `docs/research/17-search-surface-and-cross-features.md`.
+
+## Decision
+
+**1. One write path, three read products.**
+
+A brain remains the unique source of truth. Ingest (`POST /ingest/`, structured ingest, observations) writes chunks, embeddings, event graph, topics, and (after this work) a real lexical index **once**. Read products are separate HTTP contracts:
+
+| Product | Canonical API | Response shape | Default SLO | Benchmarks |
+| --- | --- | --- | --- | --- |
+| Memory | `POST /retrieve/context` | `text_context`, `triples`, passages, topics, paths | p50 < 1000 ms; no LLM | LoCoMo, LongMemEval, BEAM |
+| Search | **`POST /retrieve/search`** (new) | Ranked **hits** (id, channel, score, snippet, extras) | p50 < 200 ms excluding embed RTT; p95 budget TBD from telemetry | New `benchmarks.search` (not LoCoMo judge) |
+| Recommendations | `GET|POST /retrieve/recommend` | Ranked **nodes** + channel | Unchanged | `benchmarks.recsys` on `demorecsys` only |
+
+MCP / deep REST remains the expensive escalation lane (ADR-006): agentic traversal, cross-encoder, sufficiency retry, HyDE-style query expansion.
+
+**2. Do not overload `/retrieve/context` into a search engine.**
+
+Context stays a **memory assembler** (evidence for an answering model). Search stays a **ranker of addressable objects** (chunks, entities, events, optional structured records). A search client may *call* context afterwards; the APIs do not merge.
+
+**3. Share write-time indexes; isolate query-time defaults.**
+
+Shared (one index, many readers):
+
+- Vector tables + a **legal ANN index** (HNSW or `halfvec` HNSW at 3072-d; loud failure if skipped).
+- Lexical postings (`tsvector` / BM25 or equivalent) over chunk text and entity names.
+- Provenance, validity (`invalid_at`, `deprecated`), `happened_at`, topic memberships, hub-bridge table.
+
+Isolated (per product, flags on the request, defaults frozen per surface):
+
+- Fusion method and weights (RRF vs convex combination).
+- Candidate fanout, channels enabled, graph PPR, fact diversification.
+- Rerankers. Core: linear / small GBDT over already-computed scores. Plugins: SPLADE, ColBERT/PLAID, cross-encoder.
+
+`/retrieve/context` **must not** change its default ranking because search shipped. Opt-in only (`lexical=bm25` on context, or a one-way cutover after the non-regression gate).
+
+**4. Core vs plugin, same pattern as recsys.**
+
+| Layer | Owns | Analog |
+| --- | --- | --- |
+| Core | BM25 (or `tsvector` rank), dense ANN, RRF/CC fusion, filters, cheap feature ranker, search API | `/retrieve/recommend` train-free graph ranker |
+| Plugin (optional) | Learned sparse (SPLADE family), late interaction (ColBERT/PLAID), cross-encoder rerank, instruction-tuned embedders | `plugins/recsys-gnn`, `plugins/features-rec` |
+
+A missing plugin degrades search to core hybrid, never 500s the memory path.
+
+**5. Non-regression is a gate, not a hope.**
+
+Any change that touches a **shared** index or the default context path:
+
+1. Paired LoCoMo (or the current memory champion arm) McNemar vs the last accepted product/SOTA run: **no statistically significant accuracy drop**; point estimate not lower. Prefer retrieval-side **passage EvR held** (deterministic) over judge accuracy.
+2. Context p50 **not worse** than the post-dossier baseline (~1.4 s) by more than measurement noise; the sub-second target remains ADR-006’s, not this ADR’s.
+3. Graph EvR agreement gate from `00` (≥95% identical session sets) still applies if seeds/ANN change.
+4. Recsys HitRate on `demorecsys` unchanged unless the change is explicitly a recs feature.
+5. Search and recs eval **never** ingest into or wipe `locomoconv26*`, `beam*clean`, or other memory brains.
+
+Search quality is **not** judged by LoCoMo LLM-as-judge. It uses ranking metrics (Recall@k, nDCG@k, MRR) on a dedicated query set.
+
+**6. Cross-features are allowed when they are indexes or cheap scores, not when they are product-specific loops.**
+
+Allowed to help more than one product: BM25, HNSW, batched embeddings, metadata filters, topic coarse-to-fine, recency/validity features, a linear blend of dense+lexical+PPR.
+
+Forbidden on the memory hot path, allowed on search-deep or plugins: per-query LLM (HyDE), cross-encoder over large K, ANCE training loops, ColBERT serving.
+
+## Alternatives considered
+
+### Implement search by adding flags to `/retrieve/context`
+- Pros: one endpoint; less API surface.
+- Cons: prompt blob ≠ hit list; memory SLO and search SLO conflict; LoCoMo harness would pick up search ranking noise.
+- Rejected.
+
+### Separate search database (Elastic/Meilisearch) as source of truth
+- Pros: mature BM25.
+- Cons: two write paths, drift from the event graph, contradicts “unique source of truth.”
+- Rejected as the system of record. An **adapter** that *projects* brain indexes into an external engine is a plugin, not a second brain.
+
+### Put a distilled cross-encoder on every search and context query
+- Pros: largest precision jump in the MS MARCO literature.
+- Cons: ADR-006 already rejected this on context (50–200 ms, 5–20% of budget). Search can add it as an **optional** second stage with small K.
+- Rejected as default; accepted as plugin / `rerank=cross_encoder` on search only.
+
+### Train a BrainAPI-specific bi-encoder (DPR/ANCE) before shipping BM25
+- Pros: best dense quality in-domain.
+- Cons: needs relevance labels and an index-refresh loop; LoCoMo N cannot support it; BEIR shows off-the-shelf dense is weak zero-shot without lexical backup.
+- Rejected as a prerequisite. Off-the-shelf embeddings + real BM25 first.
+
+## Consequences
+
+- New route family under `/retrieve/search*`. Existing `/retrieve/context` and `/retrieve/recommend` keep their contracts.
+- Lexical storage and ANN legality become substrate work (`03`), not a search-plugin secret.
+- `benchmarks/search/` and `benchmarks.search` in `REPORTS.json` are required before claiming search quality, mirroring recsys isolation.
+- Memory SOTA work (`12`, `13`) may **opt into** BM25∪dense RRF on context after the gate — HyperMem (`2604.08256`) already uses that hybrid — but shipping search does not force that opt-in.
+- Documentation in `00` gains workstream `17`. Implementation does not start from this ADR alone; tasks live in `docs/research/17-search-surface-and-cross-features.md`.
+- Optional query-gated personalize on search (`target`) is [ADR-008](008-query-gated-search-personalization.md). It must not change omitted-`target` ranking, default `channels=["passages"]`, or this ADR’s 200 ms SLO.
diff --git a/docs/decisions/008-query-gated-search-personalization.md b/docs/decisions/008-query-gated-search-personalization.md
new file mode 100644
index 0000000..fe17bd8
--- /dev/null
+++ b/docs/decisions/008-query-gated-search-personalization.md
@@ -0,0 +1,97 @@
+# ADR-008: Query-Gated Personalized Search on Retrieved Hits
+
+## Status
+
+Accepted — 2026-08-21. Does not supersede ADR-007. Search remains a separate product from recommendations. This ADR locks how identity and taste may touch `/retrieve/search` without sharing recs request semantics or becoming the default ranker.
+
+Locked resolutions:
+
+1. Identity on search is optional `target` (same name as `/retrieve/recommend`). The field is never `user_id`.
+2. First-stage stays query → passage (omitted `channels=["passages"]`). Personalize reorders **already retrieved** hits; it does not call `EntityRecommendRetriever.recommend()`.
+3. User-asserted `extras` remains a hard AND filter. Inferred taste is blend-only (I7 rejected).
+4. HEM / TEM / ZAM training is deferred (I6). WANDS and ESCI cannot supply query×user labels.
+5. Omitted-`target` ranking and latency stay the ADR-007 product path (p50 < 200 ms excluding embed). Personalized search is opt-in.
+
+## Date
+
+2026-08-21
+
+## Context
+
+ADR-007 shipped search as a ranked hit list with its own SLO and eval (`benchmarks.search`). Catalog work ([24-search-catalog-graph.md](../research/24-search-catalog-graph.md)) attached `SearchHit.node_id` so a client can walk the same KB. Recommendations already take `target` on `/retrieve/recommend` and are scored as next-item HitRate on `demorecsys` ([16-recsys-eval-protocol.md](../research/16-recsys-eval-protocol.md)).
+
+Three failures follow if personalize is implemented as “recs on the search path”:
+
+1. **Wrong first-stage.** Hub-walk / recommend candidate generation on `searchbenchwandsgraph` recovered Recall@50 **0.287** vs passages **0.837**. That is missing gold, not a ranking defect.
+2. **Wrong training data.** HEM (SIGIR 2017, DOI `10.1145/3077136.3080813`) and TEM (arXiv `2005.08936`) need query×user×item logs. Frozen WANDS and ESCI have no users.
+3. **Wrong filter.** Writing `extras` from inferred style AND-drops relevant hits the user did not exclude.
+
+A fourth failure is **wrong identity**: a `user_id` field on `SearchRequestBody` forks the recommend contract and invites mixing HitRate with nDCG.
+
+Literature that actually transfers is **query-dependent** personalization — ZAM (arXiv `1908.11322`) and TEM both record that personalization does not always help — not always-on user vectors and not replacing BM25∪dense.
+
+Analysis: [25-personalized-search.md](../research/25-personalized-search.md).
+
+## Decision
+
+**1. Optional `target` on search; omit means anonymous.**
+
+`SearchRequestBody.target` and `GET /retrieve/search?target=` accept a USER uuid or id. Empty / omitted → no personalize call; hit order equals the pre-ship retrieve order. Unknown user → no-op. Schema forbids `user_id`, `product_id`, `sku`, and `brand` on the search body.
+
+**2. Rerank retrieved hits; do not generate candidates.**
+
+Personalize runs after extras AND and `node_id` attach, before the final `k` cut. Catalog `mode` personalizes `k_ret`. Hits without `node_id` get pref 0 and stay in the list. All-zero prefs preserve order. Never drop an id because taste did not match.
+
+Do not call `EntityRecommendRetriever.recommend()`. Score only the given node ids by ATTR-hub overlap with `user_pref_weights`.
+
+**3. Query-gated blend, not always-on taste.**
+
+λ is 0 when the query has digits / SKU-like tokens; otherwise it falls with token count (one token high, four+ low). Blend is min-max retrieve vs pref. This is a cheap stand-in for ZAM/TEM’s “personalize only when the query is under-specified,” not a trained attention model.
+
+**4. Write taste as PREFERS; keep EVENT for dated behavior.**
+
+Interaction `options` / `attributes` emit `USER -PREFERS-> ATTR` with the same hub uuid as catalog HAS. Options are not copied onto the product catalog blob. View / cart / purchase / favorite stay EVENT wrappers with `happened_at`. Architect persists any no-event direct triple, not only static HAS.
+
+Favorite / wishlist weight **0.7**; unknown behaviors **0.2**. Search pref vectors combine write-time PREFERS with recency-decayed EVENT→ITEM→ATTR (14-day half-life) even when PREFERS exist.
+
+**5. Inferred taste is not a filter.**
+
+User-asserted `extras` stays AND. The ranker must not inject `extras={style: …}` from history.
+
+**6. Eval isolation.**
+
+Unit tests in `tests/test_search_personalize.py`. Optional skip-ingest smoke only on `searchbenchwandsgraph`. Never wipe frozen `searchbench*` quality brains, `locomoconv*`, `beam*`, or `demorecsys`. Do not quote vs frozen WANDS **0.823** / ESCI **0.500**. Do not score recsys HitRate as search nDCG. If ledgered: `claim: architecture-demo` only.
+
+## Alternatives considered
+
+### Add `user_id` on `SearchRequestBody`
+- Pros: matches ingest JSONL and recsys datasets.
+- Cons: two identity names across retrieve; invites mixing HitRate with nDCG.
+- Rejected. Reuse `target`.
+
+### Call `/retrieve/recommend` (or `EntityRecommendRetriever`) as search first-stage
+- Pros: one ranking implementation.
+- Cons: different task (next-item vs query); communities/graph first-stage already lost Recall@50 on the architecture-demo brain.
+- Rejected.
+
+### Infer `extras` from PREFERS / recent events
+- Pros: “personalize” looks like faceted search.
+- Cons: AND-drops relevant hits the caller did not filter; confuses asserted constraints with taste.
+- Rejected (I7).
+
+### Train HEM / TEM / ZAM before shipping a personalize knob
+- Pros: literature SOTA for personalized product search.
+- Cons: no query×user logs on WANDS/ESCI; would not be the 200 ms default; blocks the opt-in rerank that the graph already supports.
+- Deferred (I6). Not a prerequisite.
+
+### Always-on personalize whenever `target` is present
+- Pros: simpler than λ.
+- Cons: ZAM/TEM evidence that specific queries get worse; navigational / SKU queries must stay retrieve-order.
+- Rejected.
+
+## Consequences
+
+- `SearchHitScores.personalize` is additive; omitted-`target` responses keep `personalize` unset.
+- Core library: `src/core/search/personalize.py`. Recommend candidate generation is unchanged aside from I4 weights.
+- Recommend `_attribute_pref_targets` still skips EVENT walks when PREFERS exist; search personalize does not. That split is intentional until recs explicitly adopt the combine.
+- Research record: `docs/research/25-personalized-search.md`. ADR-007 SLO and default channels stay the omitted-`target` path.
diff --git a/docs/research/00-scope-and-constraints.md b/docs/research/00-scope-and-constraints.md
index 6dd8736..cc39012 100644
--- a/docs/research/00-scope-and-constraints.md
+++ b/docs/research/00-scope-and-constraints.md
@@ -40,6 +40,15 @@ Anything expensive must be either moved to write time or exposed only on the dee
| `13-locomo-93-research-levers.md` | Evidence-backed research levers for HyperMem-class ≥93% (harness-first) |
| `15-ecommerce-gnn-recsys-landscape.md` | External GNN/e-commerce RecSys curriculum & industry index; BrainAPI mapping |
| `16-recsys-eval-protocol.md` | Held-out next-item eval on demorecsys via structured ingest + /retrieve/recommend |
+| `17-search-surface-and-cross-features.md` | Search as a third product surface on the same KB; shared indexes; non-regression vs LoCoMo/BEAM/recsys |
+| `18-search-eval-protocol.md` | Hybrid BM25+dense search eval on `searchbench*`; ledger `benchmarks.search`; no LoCoMo judge |
+| `19-search-esci-quality.md` | ESCI US first-stage / rerank quality track; frozen `searchbenchesci74` |
+| `20-esci-sota-targets.md` | External ESCI / commerce-search SOTA numbers; what we may and may not quote |
+| `21-esci-ce-graded-and-first-stage.md` | Graded CE and first-stage plugin arms; not the ADR-007 200 ms default |
+| `22-multilingual-ecommerce-search.md` | Multilingual first-stage (ESCI ES/JP; Italian smoke, not ESCI) and production commerce plan; `/retrieve/search` only |
+| `23-search-wands-quality.md` | Frozen WANDS passages control (nDCG@10 **0.823**); not a catalog-graph claim |
+| `24-search-catalog-graph.md` | Opt-in catalog graph + `node_id` sidecar; architecture-demo on `searchbenchwandsgraph` |
+| `25-personalized-search.md` | Query-gated rerank of already retrieved hits via optional `target`; ADR-008 |
## Decisions already taken during research
diff --git a/docs/research/16-recsys-eval-protocol.md b/docs/research/16-recsys-eval-protocol.md
index 87f2270..d1cb27a 100644
--- a/docs/research/16-recsys-eval-protocol.md
+++ b/docs/research/16-recsys-eval-protocol.md
@@ -14,7 +14,7 @@ Freshness: each structured ingest changes graph walks / attribute prefs immediat
**Not** LoCoMo / LongMemEval / BEAM. Never write to or wipe `beam1m1clean`, `locomoconv26*`, or other memory-eval brains. Ledger upserts go only to `benchmarks.recsys` in [`benchmarks/REPORTS.json`](../../benchmarks/REPORTS.json).
-Related: [ADR-002](../decisions/002-structured-ingestion-specific-processing.md), [15-ecommerce-gnn-recsys-landscape.md](15-ecommerce-gnn-recsys-landscape.md).
+Related: [ADR-002](../decisions/002-structured-ingestion-specific-processing.md), [15-ecommerce-gnn-recsys-landscape.md](15-ecommerce-gnn-recsys-landscape.md), [ADR-008](../decisions/008-query-gated-search-personalization.md) (search `target` is not this protocol).
---
@@ -100,3 +100,4 @@ Requires `BRAINPAT_TOKEN` in `benchmarks/.env`. Default brain: `demorecsys`. Rep
- Default harness backend is **graph** (`GET /retrieve/recommend`); use `--backend lightgcn` only for the offline comparator.
- Do **not** mutate `benchmarks.locomo` / `beam` / `longmemeval` ledger rows.
- Do **not** point `--brain` at memory-eval brains.
+- Do **not** score recsys HitRate as search nDCG. Optional `/retrieve/search?target=` is [25-personalized-search.md](25-personalized-search.md) / ADR-008 on `searchbench*` only, never `demorecsys`.
diff --git a/docs/research/17-search-surface-and-cross-features.md b/docs/research/17-search-surface-and-cross-features.md
new file mode 100644
index 0000000..3792cf7
--- /dev/null
+++ b/docs/research/17-search-surface-and-cross-features.md
@@ -0,0 +1,595 @@
+# 17 — Search surface on the shared knowledge base
+
+Workstream: make BrainAPI a **unique source of truth** that already powers memory and recommendations, and can also power **search**, with shared write-time features and isolated query contracts. Binding: `00-scope-and-constraints.md`, ADR-005, ADR-006, **ADR-007** (Accepted). Recs isolation: `16-recsys-eval-protocol.md`. Memory SOTA: `08`, `12`, `13`. Substrate indexes: `03` §G.
+
+**Plan mode.** This document does not change `src/` or harnesses. Tasks are sized for later implementers.
+
+Skills applied: scientific-brainstorming (ideas labeled, no automatic winner), scientific-critical-thinking (severity, construct validity), planning-and-task-breakdown (vertical slices, gates), paper-lookup (arXiv 2026-08-18).
+
+---
+
+## 0. Scope of this session
+
+**Focal question.** How can one brain serve (a) agent memory, (b) document/entity search, and (c) recommendations as *separate products* that share indexes and cheap scores, without making LoCoMo / LongMemEval / BEAM / recsys **worse**?
+
+| | Record |
+| --- | --- |
+| Purpose | Architecture + literature + implementable plan |
+| Audience | Maintainer; retrieval / substrate / eval owners |
+| Decision owner | Maintainer (ADR-007 Accepted 2026-08-18; §15 locked below) |
+| Time horizon | Core search API in weeks; plugins and LTR later |
+| In scope | Read APIs, lexical+ANN indexes, fusion, cheap rankers, search eval, plugin boundary, cross-features that are indexes or scores |
+| Out of scope | Replacing ingest; training a proprietary bi-encoder as a prerequisite; making `/retrieve/context` the search engine; MS MARCO as a product SLO; dual-use / biosafety (not implicated) |
+| Real constraints | ADR-006: no LLM on context path; p50 < 1000 ms target unmet (~1.4 s); judge cannot resolve <5–7 LoCoMo points; 3072-d HNSW skipped; keyword path is ILIKE |
+| Assumed | Event-leg supersession bug stays fixed going forward; existing unrepaired brains may still exist |
+| Negotiable | Whether context *later* opts into BM25; search p95 number; Postgres FTS vs external inverted-index plugin |
+| Unknown | Keyword vs paraphrase mix of future search tenants; corpus size per brain in production |
+| Prohibited | LoCoMo judge as a search metric; wiping memory eval brains; silent default changes on `/retrieve/context` |
+
+**Perspectives represented:** retrieval engineering, memory eval, recsys (existing plugin split), IR literature. **Missing:** a production search-traffic owner (no click log). Do not treat LoCoMo questions as a substitute for that mix.
+
+---
+
+## 1. Product model (decision, proposed)
+
+Three **read** products, one **write** path. Analogous to recsys (`16`): BrainAPI is the KB; heavy models are plugins.
+
+```text
+ POST /ingest* (chunks, graph, vectors, topics)
+ │
+ ┌─────────────────┼─────────────────┐
+ ▼ ▼ ▼
+ /retrieve/context /retrieve/search /retrieve/recommend
+ memory assembler ranked hits ranked items
+ LoCoMo/BEAM benchmarks.search benchmarks.recsys
+ ADR-006 SLO search SLO recs SLO
+```
+
+**What “search” is not.** It is not “call context and parse `text_context`.” Context exists to feed an answering model. Search exists to return **addressable hits** (chunk / entity / event / structured row) with scores, filters, and snippets.
+
+**What “cross-feature” is.** A structure computed at write time, or a cheap score at read time, that more than one product can consume without inheriting the other’s loops (PPR over event hubs, cross-encoder, LightGCN train, sufficiency retry).
+
+---
+
+## 2. Current state (located evidence)
+
+Anchors are the working tree as of 2026-08-18. Older line numbers in `02` / `07` may have drifted; behaviour below was re-checked in `retrieve.py`, `data.py`, `vectors.py`.
+
+### 2.1 Surfaces that exist
+
+| Surface | Role today | Search-engine analogue |
+| --- | --- | --- |
+| `POST /retrieve/context` | Memory: spaCy variants → dense node/rel seeds → event-centric facts + PPR + bridges + topics; passages via RRF(dense, keyword) | RAG retriever, not a SERP |
+| `GET /retrieve/` | `retrieve_data`: text + preferred entity types | Internal lookup |
+| `GET\|POST /retrieve/recommend` | Graph recommend (+ optional LightGCN plugin) | Recs, not lexical search |
+| `GET /retrieve/entity/synergies` | Sibling / synergy scores | Related-entities widget |
+| MCP `search_semantically` / `traverse_graph` / `search_memory` | Deep / agentic | Escalation, unbounded |
+
+### 2.2 The “hybrid” passage channel is not BM25∪dense
+
+```1680:1722:src/services/api/controllers/retrieve.py
+def _retrieve_passages(...):
+ ...
+ vector_hits = vector_search.search_data(...)
+ ...
+ search_result = data_adapter.search(text, brain_id) # ILIKE
+ fused = reciprocal_rank_fusion([vector_ids, keyword_ids])
+```
+
+`data.py` `search` is `WHERE text ILIKE %s` with `'%' + escaped_query + '%'`. No inverted index, no term weights, no `k1`/`b`. RRF (`fact_filter.py`, `k=60`) fuses ANN order with a **substring hit list**.
+
+**Implication (interpretation):** ADR-005’s hybrid design is real as *architecture*; the sparse leg is not the IR method the SOTA brief and HyperMem call BM25. Keyword/navigational search is the first thing this stack cannot honestly sell.
+
+### 2.3 Dense ANN is optional / silently off on default pgvector
+
+`_vector_index_ddl` returns empty string when `dimension > 2000`. Deploy examples set all stores to **3072**. `search_vectors` still `ORDER BY embeddings <=> $1` and sets `hnsw.ef_search` (no-op without an index). Over-fetch + uuid tie-break (`ann_overfetch_k`, `stable_top_k_vectors`) is in place for **determinism**, not speed. Milvus uses `AUTOINDEX` + COSINE **if** that backend is selected.
+
+On LoCoMo-sized brains this is not the latency bottleneck (`06`: passages p50 829 ms of which **682 ms embed**). On BEAM-scale or multi-million-chunk brains it becomes a sequential scan. **Two different products, two different bottlenecks.**
+
+### 2.4 Memory path latency and quality (do not mix with search claims)
+
+| Signal | Status | Use for search work |
+| --- | --- | --- |
+| Context p50 | ~1.3–1.4 s after dossier deletion; ADR-006 target 1 s | Shared embed batching helps both; do not put CE on this path |
+| Passage EvR | ~97% conv-26 | LoCoMo **cannot** show BM25 recall gains in aggregate; need a keyword slice or a non-LoCoMo corpus |
+| Graph vs passages-only judge acc | McNemar ns at n=152 | Do not justify search graph-channel by LoCoMo accuracy |
+| Graph EvR | C3: 47.3% → 56.0% with reserved bridge slots + ANN seed stability | Graph as a **search channel** (return events/entities as hits) is still a product idea, not a LoCoMo win |
+| SOTA compose arm | conv-26 95.4% judge (`13`); full-10 still open | **Do not touch default context ranking** while that claim is live unless the gate in §8 passes |
+| Recsys | Isolated brain `demorecsys`; plugins for GNN / attribute prefs | Copy this isolation for search eval |
+
+### 2.5 Ranking on context (no LTR, no CE)
+
+Distance (min over event nodes) → optional PPR **replaces** distance → MMR diversify. LLM fact filter is a no-op without a deep-tier adapter. No LambdaMART, no linear blend of BM25+dense+PPR. Historical context still re-runs the same passage search (~91.5% duplicate; cost, not wall-clock). Passage collection still **serial** over 3–6 queries each calling `embed_text` though `embed_texts` exists.
+
+### 2.6 Recs already show the split we want for search
+
+`16`: core = train-free graph ranker on `/retrieve/recommend`; `plugins/features-rec` writes extra edges; `plugins/recsys-gnn` is optional LightGCN and **does not replace** ingest. Search should copy that: core BM25+dense+filters; plugins for SPLADE / ColBERT / CE.
+
+---
+
+## 3. Literature (located, 2026-08-18)
+
+**Retrieval summary**
+
+- Query: canonical IR + memory papers by arXiv ID, plus title search for hybrid/RRF/SPLADE/ANCE
+- Scope: targeted lookup (not exhaustive)
+- Databases: arXiv Atom API (`export.arxiv.org/api/query`) parsed with `paper-lookup/scripts/arxiv_atom.py`
+- Access date: 2026-08-18
+- Limits: first page of title-OR search (`total_results=59`, retrieved 8); ID lists complete (12 + 4). RRF’s original SIGIR 2009 paper is **not on arXiv** (Cormack, Clarke, Buettcher) — cited by name only.
+
+**Warnings:** Absence from this bound is not “never studied.” UniIR (`2311.17136`) is **multimodal** instruction retrieval — weak fit for text-only BrainAPI v1. HyDE (`2212.10496`) uses an LLM at query time — incompatible with ADR-006 on context; plugin/deep only.
+
+### 3.1 What actually supports a BrainAPI search product
+
+| Paper | arXiv | Status for us | Takeaway (not a BrainAPI result) |
+| --- | --- | --- | --- |
+| BEIR | [2104.08663](https://arxiv.org/abs/2104.08663) | **support-located** (zero-shot IR) | BM25 is a strong OOD baseline; dense often loses zero-shot; hybrid / CE / late-interaction win on average at higher cost. **This is the argument for a real lexical index before a custom bi-encoder.** |
+| DPR | [2004.04906](https://arxiv.org/abs/2004.04906) | support-located (mechanism) | Dual-encoder + ANN can beat BM25 **in-domain with labels**. We have hosted embeddings, not DPR training. |
+| ANCE | [2007.00808](https://arxiv.org/abs/2007.00808) | support-located (mechanism) | Hard negatives from a live ANN index close the train/test gap. High engineering cost; not a v1 gate. |
+| ColBERT / v2 / PLAID | [2004.12832](https://arxiv.org/abs/2004.12832), [2112.01488](https://arxiv.org/abs/2112.01488), [2205.09707](https://arxiv.org/abs/2205.09707) | mixed | Token-level matching approaches CE quality; PLAID claims tens of ms on GPU at 140M passages. **Index size and serving complexity → plugin**, not core. |
+| SPLADE / v2 | [2107.05720](https://arxiv.org/abs/2107.05720), [2109.10086](https://arxiv.org/abs/2109.10086) | mixed | Learned sparse on inverted indexes; v2 reports large BEIR gains. Index growth and inference → plugin after BM25 exists. |
+| M3-Embedding (BGE-M3) | [2402.03216](https://arxiv.org/abs/2402.03216) | **idea-adjacent** | One model, three modes: dense, sparse, multi-vector, long context (8k). Strong **cross-feature candidate** if we ever change the embedder; not a silent swap (would re-embed every brain). |
+| Matryoshka / MRL | [2205.13147](https://arxiv.org/abs/2205.13147) | support-located | Truncatable embeddings; `03` already maps this to making HNSW legal. `text-embedding-3-large` supports dimension cut natively. |
+| Fusion analysis | [2210.11934](https://arxiv.org/abs/2210.11934) | **challenge-located vs “just RRF”** | Convex combination of lexical+dense scores can beat RRF; RRF is **parameter-sensitive**. Our hardcoded `k=60` is a choice, not a law. Prefer exposing RRF **and** CC on the search API. |
+| HyperMem | [2604.08256](https://arxiv.org/abs/2604.08256) | support-located (memory) | LoCoMo SOTA report uses **topic→episode→fact + BM25∪dense RRF**. We already have topic coarse-to-fine; we lack BM25. **Same hybrid is a memory cross-feature**, opt-in on context after the gate. |
+| BEAM | [2510.27246](https://arxiv.org/abs/2510.27246) | support-located (memory scale) | Long dialogues; retrieval cost grows with N. HNSW legality matters here more than on conv-26. |
+| Filtered ANN | [2602.11443](https://arxiv.org/abs/2602.11443) | support-located (substrate) | Metadata filters + ANN; Milvus recall more stable than pgvector plans in their study. Search **filters** (type, validity, time) are first-class, not a WHERE after LIMIT. |
+| Contrastive embeddings | [2201.10005](https://arxiv.org/abs/2201.10005) | background | Off-the-shelf embeddings can be strong for semantic search. Justifies keeping a general embedder in v1. |
+| HyDE | [2212.10496](https://arxiv.org/abs/2212.10496) | **veto on context** | LLM generates a hypothetical doc then embeds. Deep/search-plugin only. |
+
+**Not treated as evidence for BrainAPI quality:** MS MARCO MRR tables in the user’s SOTA brief (analytical interpretation, different task, no reproduction here).
+
+### 3.2 What would confirm / refute transferring these papers
+
+- **Confirm BM25 value:** keyword-slice Recall@10 on a brain where ILIKE misses exact tokens that BM25 posts; latency not worse than ILIKE seqscan on a 100k-chunk fixture.
+- **Refute BM25 value:** no Recall@k lift on that slice and on LoCoMo passage EvR (already saturated).
+- **Confirm HNSW:** recall@k vs exact `ORDER BY <=>` within a declared drop (e.g. ≤2 points at k=24) **and** p95 vector stage drop on a large fixture.
+- **Refute “search will lift LoCoMo judge”:** McNemar ns after BM25-on-context opt-in — expected; do not use that as a search-ship failure.
+
+---
+
+## 4. Independent idea register
+
+Stage: `independent` then clustered. Origin: `AI-assisted` + `literature-inspired` (this session). These are **proposals**, not findings.
+
+| ID | Statement | Origin | Predicted observation | Disconfirm |
+| --- | --- | --- | --- | --- |
+| S01 | Add `POST /retrieve/search` returning scored hits, not a prompt blob | product | Search clients stop scraping `text_context` | Nobody uses it; context remains the only caller |
+| S02 | Replace ILIKE with Postgres BM25/`tsvector` (or `pg_trgm` for short codes) as **core** lexical | BEIR, HyperMem | Keyword Recall@k ↑; long-query ILIKE misses ↓ | No lift on keyword slice |
+| S03 | Legalize ANN (`halfvec` HNSW or MRL truncate ≤2000) with loud skip | `03` G, MRL | BEAM/large-brain vector p95 ↓; seed stability held via overfetch | Recall@k vs exact collapses |
+| S04 | Batch `embed_texts` once per request; concurrent dense+lexical; single query on context | ADR-006 leftover | Context p50 toward <1 s | Stage timings still show multi-RTT embed |
+| S05 | Expose fusion `rrf` and `cc` (convex combination) on search | 2210.11934 | CC wins in-domain if a tiny labelled set exists | RRF ≈ CC on our mix |
+| S06 | Cheap linear/GBDT ranker over {dense, lexical, PPR, recency, topic, validity} | ADR-006, LTR brief | Precision@10 on search set ↑; microseconds | Features collinear; no lift |
+| S07 | Search channels: `passages` \| `entities` \| `events` \| `structured` | product | Entity-name queries don’t require passage hits | Graph channel noisy; users ignore non-passage |
+| S08 | Metadata filters (label, validity, time range) applied **inside** retrieval | 2602.11443 | Filtered search correct; no post-LIMIT drop | Planner ignores filter; recall hole |
+| S09 | Opt-in `lexical=bm25` on context after search BM25 is proven | HyperMem | Memory keyword misses ↓ without judge drop | EvR or McNemar regresses |
+| S10 | Plugin `features-search` or `search-rerank`: CE on K≤10 | brief, ADR-006 | Search nDCG ↑; context untouched | p95 blows search SLO |
+| S11 | Plugin late-interaction (ColBERT/PLAID) or SPLADE | ColBERT family, SPLADE | Hard queries ↑ | Index size / reindex cost dominates per-brain |
+| S12 | Optional embedder swap to BGE-M3 for dense+sparse+colbert from one model | 2402.03216 | One write, three retrieval modes | Re-embed cost; LoCoMo seed shift |
+| S13 | Recs cold-start: `/retrieve/search` on item text for new SKUs, then graph recs | recs×search | Cold items appear before walks exist | Duplicate ranking; worse HitRate |
+| S14 | Do **not** train ANCE/DPR in v1 | negative control | — | If we later get labels and a refresh loop, reopen |
+| S15 | HyDE / query LLM rewrite only on deep search | 2212.10496, ADR-006 | Hard paraphrases ↑ at high latency | Context accidentally calls it |
+| S16 | Search eval harness on a **separate** brain; ledger `benchmarks.search` | `16` analog | Claims don’t contaminate LoCoMo | People still quote judge % for search |
+
+**Minority / preserved:** S12 (embedder swap) and S11 (ColBERT core) are attractive and **wrong as v1** if they force re-ingest of memory brains. S07 graph-as-hits may look unused until a tenant searches people/events rather than documents — keep the channel even if default k focuses on passages.
+
+**What became less obvious after evidence check:** “Ship hybrid” is already the *shape* of context; the gap is the **sparse implementation** and the **missing hit-list API**, not the absence of RRF. “Need IVF+PQ / 100M HNSW RAM” is the SOTA brief’s catalog assumption; BrainAPI shards **per brain**. “Cross-encoder on K=10 for 200 ms p99” fights ADR-006 on memory and is optional on search.
+
+---
+
+## 5. Clusters (by mechanism, not by slogan)
+
+| Cluster | IDs | Shared mechanism |
+| --- | --- | --- |
+| C-API | S01, S07, S08 | New contract: hits, channels, filters |
+| C-LEX | S02, S09 | Real inverted/lexical index |
+| C-ANN | S03, S04 | Vectors actually indexed; embed RTT collapsed |
+| C-FUSE | S05, S06 | How scores combine (RRF/CC/LTR) |
+| C-PLUG | S10, S11, S12, S15 | Heavy IR off the core path |
+| C-X | S09, S13 | Cross-product: memory opt-in BM25; recs cold-start via search |
+| C-EVAL | S16 | Isolated measurement |
+| C-NOT | S14 | Explicit non-goals for v1 |
+
+Merges: S02 and “SPLADE in core” stay **split** (different index ops and failure modes). S04 is not the same as S03 (latency vs scale).
+
+---
+
+## 6. Criteria (declared before preference)
+
+Direction: higher is better unless noted. Weights are decision aids, not a computed winner.
+
+| Criterion | Weight | Anchors (1–5) | Evidence needed |
+| --- | --- | --- | --- |
+| Memory non-regression | **Gate** (noncompensatory) | Fail = LoCoMo/BEAM/recsys worse | Paired McNemar, passage EvR, recs HitRate, p50 |
+| Search usefulness | 3 | 1 = still ILIKE+blob; 5 = hit API + BM25 + filters | Recall@k / nDCG on search set |
+| Latency / SLO fit | 3 | 1 = extra LLM; 5 = index lookup + one embed | `profile_stages` |
+| Cross-feature leverage | 2 | 1 = search-only fork; 5 = one index, three readers | Same DDL used by context/search/recs |
+| Feasibility / reversibility | 2 | 1 = re-embed all brains; 5 = additive index + flag | Rollback to ILIKE/exact scan |
+| Information gain if null | 1 | Null still teaches query mix | Pre-registered slices |
+
+**Vetoes (not averaged):** LLM on `/retrieve/context`; default context ranker change without §8 gate; search eval writing memory brains; dual source of truth.
+
+---
+
+## 7. Adversarial review (short)
+
+| Idea | Wrong if… | Alternative explanation | Hidden dependency | Harm |
+| --- | --- | --- | --- | --- |
+| S02 BM25 | LoCoMo-like questions are long paraphrases; BM25 adds posting I/O and zero recall | Dense already covers; ILIKE was enough on tiny N | Tokenization mismatch with embedder tokenizer | Over-rank boilerplate terms |
+| S03 HNSW | Approx top-k reshuffles seeds → graph EvR jitter (C3 already fought this) | Latency win is still embed RTT | `halfvec` ops / pgvector version | Silent quality flake on memory |
+| S01 new API | All “search” users were agents who wanted context anyway | Product confusion; two ways to do one thing | Client SDK + docs | Split brain of integrations |
+| S06 LTR | No labels except LoCoMo sessions (biased to memory) | Overfit to session IDs | Click log does not exist | Ranking that looks good offline, worse UX |
+| S12 M3 embedder | Different geometry vs current 3072-d OpenAI vectors | “One model three modes” marketing | Full re-embed | SOTA compose arm incomparable |
+
+**Anchoring risk:** the user’s SOTA brief pushes hybrid→CE→LTR on a 200 ms p99 web SLO. We keep the **sequence of primitives** (lexical, ANN, optional CE) but **reject the SLO and the default CE**.
+
+---
+
+## 8. Non-regression protocol (how we refuse to make memory worse)
+
+Copy recsys isolation (`16`, `benchmarks/AGENTS.md`).
+
+### 8.1 What may change without a memory A/B
+
+- New routes under `/retrieve/search*` that do not alter context defaults.
+- New tables/indexes that context **does not read** yet (e.g. `tsvector` column unused by `_retrieve_passages`).
+- Plugins that only register search rerank.
+
+### 8.2 What requires the gate
+
+- Context reading BM25 instead of ILIKE.
+- Enabling HNSW / changing embedding dimensionality or metric.
+- Changing RRF `k`, seed `k`, passage fanout, PPR default, fact diversify.
+- Shared embed batching that changes variant set or seed order (stabilize with existing uuid/distance quantization).
+
+### 8.3 Gate (all must hold)
+
+1. **Memory champion arm** (document the exact `run_id` at implementation time; today the conv-26 SOTA pointer is `locomo-compose-sota-conv26-v4d` in `REPORTS.json`): paired McNemar **not** significantly worse; point estimate ≥ baseline. Prefer **product** profile for default-path changes; do not require SOTA compose (SC/gap-fill) to move.
+2. Passage **full evidence-session recall** held to the digit on that brain (deterministic). Graph EvR: if ANN/seeds change, identical-config agreement ≥95% (`00`).
+3. Context p50: not worse than the then-current post-dossier baseline beyond noise (record both numbers). Sub-second remains ADR-006, not a search ship blocker.
+4. Recsys: `demorecsys` HitRate@K unchanged unless the PR is a recs PR.
+5. Brains: search fixtures use IDs like `searchbench*`. Never wipe `locomoconv26*`, `beam*clean`, `demorecsys`.
+
+**If BM25-on-context is null on LoCoMo:** keep ILIKE as context default; search still ships BM25. That is a successful split, not a failed hybrid.
+
+---
+
+## 9. Proposed search API (idea → later spec)
+
+Vertical slice: one call that a UI or engine can use without parsing memory blobs.
+
+```http
+POST /retrieve/search
+```
+
+```json
+{
+ "query": "alice counseling license",
+ "brain_id": "default",
+ "k": 20,
+ "channels": ["passages", "entities"],
+ "lexical": "bm25",
+ "dense": true,
+ "fusion": "rrf",
+ "filters": {
+ "labels": ["PERSON", "EVENT"],
+ "currently_valid": true,
+ "happened_after": null
+ },
+ "rerank": "none",
+ "profile_stages": false
+}
+```
+
+```json
+{
+ "hits": [
+ {
+ "id": "chunk_…",
+ "channel": "passages",
+ "score": 0.031,
+ "scores": { "bm25": 12.4, "dense": 0.22, "rrf": 0.031 },
+ "snippet": "…",
+ "source_session_ids": ["session_3"],
+ "entity_uuid": null
+ }
+ ],
+ "insufficient": false,
+ "stage_timings": null
+}
+```
+
+Optional later: `GET /retrieve/search/suggest` (prefix / trgm). Out of v1 if it expands scope.
+
+`rerank`: `none` | `linear` | `plugin:<name>`. Unknown plugin → 400, not a silent fallback that looks like core quality.
+
+---
+
+## 10. Cross-feature map
+
+| Feature | Memory | Search | Recs | When to share |
+| --- | --- | --- | --- | --- |
+| BM25 / tsvector | Opt-in passage channel (HyperMem-like) | **Default lexical** | Item-text cold start (S13) | Write once; read per product flag |
+| HNSW / halfvec | Seed + passage ANN at scale | Same | Node similarity if used | Shared; gate on seed stability |
+| Batched embed | Closes ~400 ms gap (`06`) | One RTT per search query | If recs embed queries | Shared helper, no rank change if vectors identical |
+| Topic memberships | Coarse-to-fine already on context | Facet / boost | Weak | Already written |
+| Validity / `happened_at` | Current-truth vs history (`05`) | Filters | Recency decay (already on recs) | Shared properties |
+| PPR / hub bridges | Multi-hop coverage | `events` channel ranking | Not default | **Do not** turn search into a second context assembler |
+| Linear LTR features | ADR-006 context ranker | Search precision | Could consume overlap scores | Train **per product** or freeze a tiny global blend with separate weights |
+| Cross-encoder | Deep tier only | Search plugin, K≤10 | No | Plugin |
+| LightGCN | No | No | Plugin | Unchanged |
+
+**Prediction:** the highest cross-leverage items are **S02+S03+S04** (indexes + embed batch). They improve search immediately and *can* improve memory latency/keyword recall without touching judge prompts.
+
+---
+
+## 11. Architecture decisions for implementers
+
+- **Lexical engine v1:** Postgres FTS (`tsvector` + `ts_rank_cd` or a BM25 extension available in the deployed image). Avoid requiring Elasticsearch as system of record (ADR-007). A plugin may *project* to an external engine later.
+- **Do not change `EMBEDDING_*_DIMENSION` as a drive-by** in the search PR. HNSW legality is a dedicated slice with a recall-vs-latency measurement (`00`, `03`).
+- **Context keeps ILIKE** until S09 gate. Implementation may write tsvector in the same migration.
+- **RRF stays** as default fusion (already in code, HyperMem-like). Expose `fusion=cc` with `alpha` for search; do not retune context `k=60` in the same PR.
+- **No ANCE/DPR training, no ColBERT core, no HyDE on context** in this workstream’s v1.
+
+---
+
+## 12. Search evaluation protocol
+
+Mirror `16` so search cannot hide inside LoCoMo. Binding write-up: [`18-search-eval-protocol.md`](18-search-eval-protocol.md). Harness: `benchmarks/search/` + `./search.sh`.
+
+| Item | Rule |
+| --- | --- |
+| Brain | `searchbench*` only |
+| Harness | `benchmarks/search/` + `./search.sh`; HTTP only |
+| Ledger | `benchmarks.search` in `REPORTS.json` only |
+| Metrics | Recall@{5,10,20}, nDCG@10, MRR; p50/p95 of `/retrieve/search` **excluding** `embed.query`; optional keyword/paraphrase slice |
+| Query mix | At least: (a) **keyword / name**, (b) **paraphrase / semantic**. Filter/time deferred until the search API accepts filters. Gold = chunk IDs via `DOCID` markers, not LLM judge |
+| Cheap start | Reuse a **held-out** subset of memory questions as *retrieval* labels (gold session → chunks via provenance) **without** running the answerer. Report as `search-from-locomo-qrels`, never as LoCoMo accuracy. Not in harness v1 |
+| Forbidden | Ingest into memory eval brains; quoting judge % as search quality |
+
+Until a scored `evaluate` run is on the `benchmarks.search` leaderboard, do not claim “search is better than BM25-in-the-brief.” Toy `search_toy` rows are isolation checks, not BEIR.
+
+---
+
+## 13. Task list
+
+Vertical slices. Each leaves the system working. No task is XL.
+
+### Phase 0 — Contract and isolation
+
+## Task 1: Freeze the three-surface contract
+
+**Description:** ADR-007 accepted or explicitly deferred by the maintainer; this doc linked from `00`. No ranking code.
+
+**Acceptance criteria:**
+- [ ] Maintainer records accept / amend / reject on ADR-007
+- [ ] `00` workstream row for `17` remains accurate
+
+**Verification:** Human review of ADR-007.
+
+**Dependencies:** None
+**Files:** `docs/decisions/007-three-product-surfaces-one-kb.md`, `docs/research/00-scope-and-constraints.md`
+**Scope:** S
+
+## Task 2: Search eval skeleton
+
+**Description:** Package `benchmarks/search/` with smoke: ingest N chunks into `searchbenchsmoke`, call a **stub** or interim `POST /retrieve/search` (may 501 until Task 4), ledger key documented in `benchmarks/AGENTS.md`.
+
+**Acceptance criteria:**
+- [ ] `./search.sh smoke` fails loudly if it would use a memory brain id
+- [ ] `REPORTS.json` schema comment or empty `benchmarks.search` leaderboard allowed
+
+**Verification:** `./search.sh smoke` documented; unit test on brain-id guard.
+
+**Dependencies:** Task 1 (or proceed as Proposed)
+**Files:** `benchmarks/search/**`, `benchmarks/AGENTS.md`
+**Scope:** M
+
+### Checkpoint A
+- [ ] Product split is written; search eval cannot wipe LoCoMo/BEAM brains
+
+---
+
+### Phase 1 — Core search first stage (does not change context defaults)
+
+## Task 3: Lexical index at write time
+
+**Description:** Persist `tsvector` (or equivalent) on `data_text_chunks` (and entity name field if cheap). GIN index. Backfill existing brains lazily or via explicit job. Context **does not** query it yet.
+
+**Acceptance criteria:**
+- [ ] New chunks get lexical vectors on ingest
+- [ ] ILIKE path still used by `_retrieve_passages`
+- [ ] Unit test: token query matches a chunk ILIKE would miss or vice versa, documented
+
+**Verification:** `pytest` on FTS helpers; explain/analyze on a fixture shows index use.
+
+**Dependencies:** None (can parallel Task 2)
+**Files:** `src/lib/postgresql/data.py`, ingest write path, tests
+**Scope:** M
+
+## Task 4: `POST /retrieve/search` core hybrid
+
+**Description:** New controller: embed once, dense `search_data` + BM25/tsvector, fusion RRF default, return hits. No PPR, no fact assembler, no historical duplicate.
+
+**Acceptance criteria:**
+- [ ] OpenAPI/schema: query, k, channels⊇passages, lexical, dense, fusion
+- [ ] Does not import dossier/synergy retrievers
+- [ ] `profile_stages` works like context
+- [ ] Context tests unchanged (no default behavior change)
+
+**Verification:** `pytest` API tests; `./search.sh smoke` green against live API.
+
+**Dependencies:** Task 3
+**Files:** `src/services/api/routes/retrieve.py`, new controller module, `requests.py`
+**Scope:** M
+
+## Task 5: Filters + entity channel (optional second slice)
+
+**Description:** `currently_valid`, label filter; optional `entities` channel via node ANN + name FTS.
+
+**Acceptance criteria:**
+- [ ] Filter is applied so that invalid entities are not in top-k when `currently_valid=true`
+- [ ] Passage-only default remains if `channels` omitted
+
+**Verification:** Tests with a superseded edge / deprecated node.
+
+**Dependencies:** Task 4
+**Files:** search controller, graph/data queries
+**Scope:** M
+
+### Checkpoint B
+- [ ] Search API returns BM25∪dense hits
+- [ ] `/retrieve/context` byte-level ranking defaults unchanged (diff + LoCoMo smoke optional)
+
+---
+
+### Phase 2 — Shared speed/scale (gated)
+
+## Task 6: Batch embeddings on context **and** search
+
+**Description:** One `embed_texts` per request; concurrent dense+lexical; collapse context passage variant loop toward ADR-006 (single query or batched variants sharing the batch). Preserve seed uuid/distance stabilization.
+
+**Acceptance criteria:**
+- [ ] Stage timings: a single `embed.query` (or one batched stage) on a profiled request
+- [ ] Passage EvR held on a recorded memory arm **if** context code changed
+- [ ] Search p50 excluding network policy: one embed RTT
+
+**Verification:** `profile_stages` JSON; paired EvR if context touched.
+
+**Dependencies:** Task 4 if sharing helper; else can start from context-only
+**Files:** `retrieve.py`, embeddings adapter call sites
+**Scope:** M
+
+## Task 7: Legalize HNSW
+
+**Description:** Measurement-first: `halfvec` vs truncate-to-2000 vs Milvus-only. Loud log/metric when index DDL skipped. Keep overfetch+uuid order.
+
+**Acceptance criteria:**
+- [ ] Default 3072 install either has an ANN index **or** fails closed with a visible error at store init
+- [ ] Recall@k vs exact on a ≥10k vector fixture recorded
+- [ ] Memory gate §8 if context seeds change
+
+**Verification:** Fixture script + notes in run dir; no silent seqscan.
+
+**Dependencies:** None vs search API; **gate** before flipping production default
+**Files:** `src/lib/postgresql/vectors.py`, config, tests
+**Scope:** M
+
+### Checkpoint C
+- [ ] Context p50 moved toward 1 s **or** numbers escalate per ADR-006
+- [ ] ANN skip cannot be silent
+
+---
+
+### Phase 3 — Fusion / cheap LTR / memory opt-in
+
+## Task 8: Search fusion knobs + linear rerank
+
+**Description:** `fusion=rrf|cc` with documented `alpha`; optional `rerank=linear` using dense, lexical, recency, validity — **no LLM**.
+
+**Acceptance criteria:**
+- [ ] Default remains RRF k=60-equivalent
+- [ ] Linear weights are config/request, not a trained black box in v1
+- [ ] Search nDCG measured on `searchbench*` vs RRF-only
+
+**Verification:** `./search.sh evaluate` two arms.
+
+**Dependencies:** Task 2, Task 4
+**Files:** search controller, `fact_filter.py` or shared fusion
+**Scope:** S–M
+
+## Task 9: Opt-in BM25 on context (cross-feature)
+
+**Description:** `GetContextRequestBody.lexical: ilike | bm25` default **`ilike`**. Document HyperMem-style union as the reason to try `bm25`.
+
+**Acceptance criteria:**
+- [ ] Default path = today’s ILIKE RRF
+- [ ] `lexical=bm25` uses the Task 3 index
+- [ ] Gate §8 on a product LoCoMo arm before considering default flip
+
+**Verification:** McNemar + passage EvR; do not use SOTA compose as the only arm.
+
+**Dependencies:** Task 3, Task 6 recommended
+**Files:** `retrieve.py` `_retrieve_passages`, `requests.py`
+**Scope:** S
+
+### Checkpoint D
+- [ ] Search has fusion/linear knobs
+- [ ] Memory default still ILIKE unless gate passed and maintainer flips
+
+---
+
+### Phase 4 — Plugins (after core is real)
+
+## Task 10: Search rerank plugin hook
+
+**Description:** Same pattern as `recsys-gnn`: optional package registers `rerank=plugin:cross-encoder` (or SPLADE document expansion). Core works if plugin absent.
+
+**Acceptance criteria:**
+- [x] Plugin yaml + route or entrypoint
+- [x] Context path cannot call the plugin
+- [x] K≤10 documented for CE
+
+**Verification:** Load/unload plugin; 400 on unknown name.
+
+**Dependencies:** Task 4
+**Files:** `plugins/search-rerank/` or `features-search/`, plugin loader
+**Scope:** M
+
+## Task 11: Recs cold-start via search (optional)
+
+**Description:** When graph walks return < k items, fill from `/retrieve/search` on item text / attributes. Flag on recommend request, default off.
+
+**Acceptance criteria:**
+- [ ] Default HitRate on `demorecsys` unchanged
+- [ ] Flag-on arm recorded separately in `benchmarks.recsys`
+
+**Verification:** `./recsys.sh evaluate` both flags.
+
+**Dependencies:** Task 4
+**Files:** `recommend.py`, recsys harness
+**Scope:** S–M
+
+### Checkpoint E
+- [ ] Heavy IR is optional
+- [ ] Recs default unchanged
+
+---
+
+## 14. Risks
+
+| Risk | Impact | Mitigation |
+| --- | --- | --- |
+| HNSW changes memory seeds | High | Overfetch+uuid; EvR agreement gate; exact scan fallback flag |
+| BM25 on context looks “free SOTA” | Med | Opt-in; HyperMem is not our judge protocol (`13`) |
+| Search API unused; context still scraped | Med | SDKs/docs; chatbot plugin uses search for “find” and context for “answer” |
+| FTS tokenization ≠ embed tokenizer | Med | Document analyzer; don’t expect identical hits |
+| Scope bleed into ColBERT/ANCE | High | C-NOT / Task 14; plugins only |
+| Judge-noise used to “prove” search | High | Separate metrics and brains |
+
+---
+
+## 15. Open questions (maintainer)
+
+Locked 2026-08-18 (see ADR-007 Status):
+
+1. **Routes.** Keep search under `/retrieve` as `GET|POST /retrieve/search`. Do not replace `GET /retrieve/`.
+2. **Lexical engine.** Env-gated. When on, Okapi BM25 over Postgres `tsvector` + GIN on `pgvector/pgvector:pg16`. Do not require `pg_search` / `rum`. Do not ship `ILIKE` or `ts_rank_cd` as the search lexical ranker.
+3. **3072-d ANN.** Keep float32 vectors. `halfvec` HNSW (≤4000-d) + exact float32 cosine rerank of overfetch. Do not truncate / MRL-cut dimensions.
+4. **Chatbot plugin.** Out of scope for v1. No `plugins/chatbot-memory` client.
+5. **SLO.** p50 **< 200 ms excluding embed RTT**. `profile_stages` must expose `embed.query` separately from retrieve.
+6. **Context vs search ranking.** Search default is both channels fused (`SEARCH_FUSION=rrf`). Context stays today’s dense ∪ ILIKE unless `SEARCH_ENABLED=true` and `CONTEXT_PASSAGE_MODE` is left at `hybrid` (or set to `bm25` / `dense`). `ilike` freezes the memory eval path.
+
+---
+
+## 16. Overall assessment
+
+BrainAPI is already a **knowledge base** with a strong memory assembler and a recs surface that knows how to keep plugins off the default path. It is **not** yet a search engine: there is no hit-list API, no BM25, and default vectors may be unindexed.
+
+The SOTA search brief is useful as a **checklist of primitives** (lexical, dense ANN, optional CE, LTR). It is the wrong **product recipe** if pasted onto `/retrieve/context` (wrong SLO, wrong response shape, wrong eval).
+
+Highest-leverage path that can help search *and* memory without betting LoCoMo: **write a real lexical index and a legal ANN index, batch embeddings, ship `/retrieve/search` that reads them, keep context defaults frozen until a paired gate.** Cross-encoders, SPLADE, ColBERT, HyDE, and ANCE stay plugins or deep-tier — the same split recsys already uses.
+
+**Next action:** maintainer accept/amend ADR-007, then Task 2+3 in parallel (eval skeleton + tsvector), then Task 4 (search API). Label: **protocol development**, not a validated quality claim.
+
+**Status note (2026-08-21):** ADR-007 accepted; `/retrieve/search` shipped. Subsequent workstreams: `18`–`24` (eval, ESCI, WANDS, catalog graph), `25` + [ADR-008](../decisions/008-query-gated-search-personalization.md) (query-gated personalize). §16 above is the 2026-08-18 assessment that motivated the ship. Default omitted `channels=["passages"]` and the 200 ms SLO are unchanged.
diff --git a/docs/research/18-search-eval-protocol.md b/docs/research/18-search-eval-protocol.md
new file mode 100644
index 0000000..199401e
--- /dev/null
+++ b/docs/research/18-search-eval-protocol.md
@@ -0,0 +1,189 @@
+# Search eval protocol — hybrid BM25 + dense (`/retrieve/search`)
+
+How to ingest a labeled toy corpus into BrainAPI and score ranked hits without contaminating memory eval.
+
+**Role split (locked):**
+
+| Component | Responsibility |
+| --- | --- |
+| BrainAPI | KB + **core hybrid search**: `POST /ingest/` writes chunks (tsvector + dense); `POST /retrieve/search` returns ranked hits (`fusion=rrf` default, `cc` override). `rerank=none` (default). Unknown `plugin:<name>` is **400**, never a silent fallback |
+| `plugins/search-rerank` | Optional second-stage cross-encoder via `rerank=plugin:cross-encoder`. Default `mode=default` reranks at most `RERANK_MAX_K=10`. Opt-in `mode=catalog` retrieves `k_ret=min(200, max(k, 50))` and reranks at most `CATALOG_RERANK_MAX_K=50`, then cuts to request `k`. Health: `GET /search-rerank/health`. Must not run on `/retrieve/context`. Catalog+CE is **not** the ADR-007 200 ms default path |
+| `plugins/search-splade` | Learned-sparse first stage. Own index: `POST /search-splade/index`. Retrieve via `channels=["plugin:splade"]` (optionally fused with `passages`) |
+| `plugins/search-colbert` | Late-interaction first stage (MaxSim). Own index: `POST /search-colbert/index`. Retrieve via `channels=["plugin:colbert"]` |
+
+**Not** LoCoMo / LongMemEval / BEAM / RecSys. Never write to or wipe `beam*`, `locomoconv*`, `demorecsys`, or LongMemEval brains. Ledger upserts go only to `benchmarks.search` in [`benchmarks/REPORTS.json`](../../benchmarks/REPORTS.json). Never quote LoCoMo judge % as a search score.
+
+Related: [ADR-007](../decisions/007-three-product-surfaces-one-kb.md), [ADR-008](../decisions/008-query-gated-search-personalization.md), [17-search-surface-and-cross-features.md](17-search-surface-and-cross-features.md) §12.
+
+---
+
+## Server prerequisites
+
+The API under test must have:
+
+- `SEARCH_ENABLED=true`
+- `DATA_DB=postgresql`
+- `BRAIN_CREATION_ALLOWED=true`
+- At least one of `SEARCH_USE_DENSE` / `SEARCH_USE_BM25` true (product default: both)
+
+Harness auth: `BRAINPAT_TOKEN` and `BRAINAPI_URL` in `benchmarks/.env`.
+
+A 404 on `/retrieve/search` is **search disabled**, not a ranking miss. The harness must fail loudly (`Set SEARCH_ENABLED=true`).
+
+---
+
+## Pipeline
+
+```text
+docs JSONL → POST /ingest/ (skip_enrichment=true → chunk + embed only)
+ → optional POST /ingest/structured (`--ingest-graph`; entity uuid = catalog `doc_id`)
+ → optional interaction JSONL (`--interactions`; EVENT + `happened_at` on `searchbench*`)
+ → GET /retrieve/text-chunks (map DOCID markers → chunk ids)
+ → POST /retrieve/search (profile_stages=true, fusion=rrf|cc;
+ channels: passages | entities | events | communities;
+ expand: none | neighbors;
+ mode: default | catalog)
+ → Recall@{5,10,20} / nDCG@10 / MRR
+ → p50/p95 retrieve excluding embed.query
+```
+
+Gold matches **chunk ids** (via DOCID mapping) **or** `hit.id == doc_id` (graph entity uuid). Binary relevance is enough for the toy fixture. Catalog datasets carry graded `gold_grades` and nDCG@10 uses those gains.
+
+Each ingested document embeds a unique marker `DOCID <doc_id>`. After ingest, paginate `GET /retrieve/text-chunks` and map `doc_id → {chunk_id, ...}` by substring. A hit matches if `hit.id` is in that gold set **or** equals the catalog `doc_id`.
+
+Search ingest defaults to `skip_enrichment=true` (chunk + embedding, no Scout/Architect). Pass `--enrich` to restore full KG ingest. Catalog evals should stay on the skip path so OpenAI/DeepSeek extract quota is not in the loop. Embeddings still run. `--ingest-graph` writes generic HAS triples (entity + CLASS/TYPE/ATTR hubs) via product `catalog_graph`. Core search has no PRODUCT/CATEGORY/BRAND field names; those strings may appear only as ordinary node types in harness JSONL.
+
+---
+
+## Query mix
+
+Bundled fixture: [`benchmarks/data/search_toy.jsonl`](../../benchmarks/data/search_toy.jsonl).
+
+| Slice | Intent |
+| --- | --- |
+| `keyword` | Name / token overlap (BM25 should help) |
+| `paraphrase` | Semantic rewrite (dense should help) |
+
+Product-search corpora (download, not committed):
+
+| Dataset | Command | Default slice | Gains |
+| --- | --- | --- | --- |
+| Amazon ESCI / Shopping Queries (KDD Cup 2022) | `./search.sh download --name esci` | US, Task 1 `small_version`, `test` | E=1, S=0.1, C=0.01, I=0 |
+| WANDS (Wayfair) | `./search.sh download --name wands` | `query_class` | Exact=1, Partial=0.5, Irrelevant=0 |
+
+Defaults cap the local JSONL at 80 queries / 2000 products / 40 candidates per query so ingest is feasible. Frozen WANDS quality slice is `data/search_wands.jsonl` (caps 80/2000/40; actual n may be lower because `_select_catalog` drops queries with no remaining gold once `max_docs` fills). `download --name wands` refuses to overwrite that file once it exists. Use `--brain searchbenchesci` / `searchbenchwands`. Do **not** reuse `demorecsys`. Structured recsys JSONL is not a search qrel. WANDS first-stage control is passages-only, k=50, `rerank=none` — see [23-search-wands-quality.md](23-search-wands-quality.md). Graph `--ingest-graph --channels passages,entities,communities` is not the quality control. Opt-in catalog graph (hub intersection + node hybrid) uses a **new** brain `searchbenchwandsgraph` and is an **architecture demo** — see [24-search-catalog-graph.md](24-search-catalog-graph.md). Do **not** mix those numbers with frozen WANDS 0.823 or ESCI 0.500. `--ingest-graph` / `--interactions` on `searchbenchwands`, `searchbenchesci74`, `searchbenchescies`, or `searchbenchesciltr2` is refused.
+
+Filter / time slices are **deferred** until `/retrieve/search` accepts filters. Do not fake them with ILIKE on the harness side.
+
+Graph channels (core, use-case agnostic): `entities` (node ANN ∪ name CONTAINS), `events` (EVENT nodes; empty-ok on catalog-only brains; recency from `happened_at` when present), `communities` (typed hub nodes, default labels `TYPE,CLASS,TOPIC` via `SEARCH_COMMUNITY_LABELS`), `expand=neighbors` (depth-1, fanout cap + degree IDF). Default omitted `channels` remains `passages`. Hits are a mixed list of chunk ids and node uuids until a join table exists. Communities are **not** Leiden / `kg_topic_sessions`. Do not score `/retrieve/recommend` HitRate as search.
+
+Optional later arm: reuse held-out memory questions as *retrieval* labels (`search-from-locomo-qrels`) **without** running the answerer, and never as LoCoMo accuracy. Not in v1 of this harness.
+
+---
+
+## Metrics and SLO
+
+| Metric | Definition |
+| --- | --- |
+| Recall@k | \|retrieved ∩ gold\| / \|gold\| at k ∈ {5,10,20} |
+| nDCG@10 | Binary on toy gold; graded `gold_grades` on ESCI/WANDS. Hits collapse to `doc_id` (chunk map **or** node uuid) |
+| MRR | Reciprocal rank of the first gold hit |
+| `p50_retrieve_ms` / `p95_retrieve_ms` | Prefer `stage_timings` `search.retrieve.wall_ms`; else `client_wall − embed.query.wall_ms` |
+
+ADR-007 SLO: **p50 < 200 ms excluding embed RTT**. Record `embed.query` separately so the ledger cannot mix the two. That SLO is the **default** path (`mode=default`, `rerank=none`). `mode=catalog` plus `rerank=plugin:cross-encoder` is an opt-in two-stage; do not cite its client wall as the 200 ms product claim.
+
+Headline for the ledger: `ndcg@10`, `recall@10`, `p50_retrieve_ms`. Label `dataset` (`search_toy.jsonl` vs `search_esci.jsonl` / `search_wands.jsonl`) so a toy run is not mistaken for product search.
+
+A passing **smoke** is a health check, not a published quality claim. Only `evaluate` with `status: ok` upserts `benchmarks.search`.
+
+---
+
+## Harness
+
+HTTP only (`benchmarks/search/` must not import `src/`, except `search/mapping.py` which re-exports `src.core.search.catalog_graph` after putting the repo root on `sys.path`). Default brain: `searchbenchsmoke`. Any `--brain` **must** start with `searchbench`.
+
+```bash
+cd benchmarks
+./search.sh dataset-stats
+./search.sh smoke
+./search.sh evaluate --fusion rrf
+./search.sh evaluate --fusion cc --run search-toy-cc
+./search.sh download --name esci
+./search.sh download --name esci --locale es
+./search.sh download --name wands
+./search.sh --brain searchbenchitsmoke evaluate --dataset data/search_italian_smoke.jsonl --run search-italian-smoke
+./search.sh evaluate --dataset data/search_esci.jsonl --brain searchbenchesci --run search-esci
+./search.sh --brain searchbenchwands evaluate --dataset data/search_wands.jsonl --run search-wands-passages-k50 --channels passages --k 50 --ks 5,10,20,50
+./search.sh --brain searchbenchwands evaluate --dataset data/search_wands.jsonl --run search-wands-passages-k50 --channels passages --k 50 --ks 5,10,20,50 --skip-ingest
+./search.sh --brain searchbenchwandsgraph evaluate --dataset data/search_wands.jsonl --run search-wandsgraph-communities-k50 --channels communities --k 50 --ks 5,10,20,50 --ingest-graph
+./search.sh evaluate --rerank plugin:cross-encoder --run search-toy-ce
+./search.sh evaluate --mode catalog --rerank plugin:cross-encoder --k 50 --brain searchbenchesci74 --skip-ingest --dataset data/search_esci_74.jsonl --run search-esci-74-catalog-ce-k50
+./search.sh rank-corpus --dataset data/search_esci_74.jsonl --run search-esci-74-exhaustive-ce
+./search.sh list-overlap --passages-run search-esci-74-passages-k50 --against-runs search-esci-74-bge-base-k50,search-esci-74-colbert-k50
+./search.sh union-lists --from-runs search-esci-74-passages-k50,search-esci-74-bge-base-k50,search-esci-74-colbert-k50 --run search-esci-74-union-bge-k50
+./search.sh cascade-lists --passages-run search-esci-74-passages-k50 --from-runs search-esci-74-bge-base-k50,search-esci-74-colbert-k50 --run search-esci-74-cascade-tail-k50
+./search.sh finetune-4class --dataset data/search_esci_74.jsonl --base microsoft/deberta-v3-base --epochs 2 --max-pairs 80000 --batch-size 8 --out data/models/esci-deberta-v3-base-4class
+SEARCH_RERANK_MODEL=data/models/esci-deberta-v3-base-4class ./search.sh rerank-retrieved --from-run search-esci-74-passages-k50 --dataset data/search_esci_74.jsonl --run search-esci-74-passages-k50-ce-deberta --ks 5,10,20,50
+./search.sh evaluate --channels passages,plugin:splade --run search-toy-splade
+./search.sh smoke --interactions data/recsys_toy.jsonl --channels events --brain searchbenchevents
+./search.sh report --run <run_id>
+```
+
+Requires `BRAINPAT_TOKEN` in `benchmarks/.env`.
+
+**Live stack note:** TUI / `brainapi start` may run from `~/.brainapi/source`. Sync or restart from the intended checkout before measuring, or scores will reflect stale code (`SEARCH_ENABLED` must be on in **that** process).
+
+---
+
+## Guardrails
+
+- Score only via `POST /retrieve/search`. Do **not** call `/retrieve/context` as the metric path.
+- Do **not** call `/retrieve/recommend`. Optional search `target` is query-gated rerank of retrieved hits (ADR-008), not this recommend protocol.
+- Do **not** mutate `benchmarks.locomo` / `beam` / `longmemeval` / `recsys` ledger rows.
+- Do **not** point `--brain` at memory-eval or recsys brains (`searchbench*` only; never `demorecsys`).
+- Failed / empty runs stay off the leaderboard.
+- Cross-encoder, SPLADE, and ColBERT are plugins. Core hybrid works if they are absent. `--rerank plugin:<missing>` and `--channels plugin:<missing>` must 400, never look like a ranking miss.
+
+---
+
+## Multilingual catalog paths (2026-08-19)
+
+**Located evidence:** `download --locale` already filtered ESCI parquet (`us` / `es` / `jp`) but defaulted to the same `search_esci.jsonl` as US. **Decision:** US stays `search_esci.jsonl`; ES/JP write `search_esci_{locale}.jsonl`; `--locale it` errors. ESCI has no Italian split (Reddy HTML; arXiv Italian product-search query empty). Fixture `data/search_italian_smoke.jsonl` is a pipeline smoke (`slice=italian-smoke`), not ESCI Task 1 and not an Italian quality claim. FTS analyzer stays `to_tsvector('english'`. Do not re-ingest `searchbenchesci74`. Plan: [22-multilingual-ecommerce-search.md](22-multilingual-ecommerce-search.md).
+
+**ES first-stage (2026-08-19, located evidence):** `./search.sh download --name esci --locale es` wrote `data/search_esci_es.jsonl` (n=62 queries, 2000 docs, `slice=esci-es`) without changing `search_esci.jsonl` / `search_esci_74.jsonl`. Live `./search.sh --brain searchbenchescies evaluate --dataset data/search_esci_es.jsonl --run search-esci-es-passages-k50 --channels passages --k 50 --ks 5,10,20,50` → nDCG@10 **0.577**, Recall@10 **0.353**, Recall@50 **0.914**, p50 84 ms. Protocol: first-stage shared-index, passages, `rerank=none`. Do not mix with US n=74 0.500/0.379 or Reddy ES 0.849 (ranking-in-pool). Product default unchanged.
+
+Italian smoke is pipeline-only (`./search.sh --brain searchbenchitsmoke evaluate --dataset data/search_italian_smoke.jsonl`). Inflected BM25 miss on english FTS; optional `SEARCH_FTS_BRAINS=searchbenchitsmoke` + `SEARCH_FTS_REGCONFIG=italian` adds `search_tsv_alt` on that brain only. `_SEARCH_DDL` stays english.
+
+**Phase 2 (2026-08-19).** Italian product-search qrels: still **no direct evidence located** (arXiv 0; OpenAlex incidental hits are not qrels). ES spelling harness `search-esci-es-spell-k50` skip-ingest: nDCG@10 0.595 (win bar 0.597) / Recall@10 0.362 / Recall@50 0.914 — **null**, not live. Optional `SearchRequestBody.extras` equality filter + `SearchResponse.facets` on hit extras; live `searchbenchitsmoke` `locale=it` keeps 3 docs, `color=nope` empty. MiniLM local-dense `search-italian-minilm-pipeline` on labeled `searchbenchitmmini` retrieved all 3 golds at k=10 (pipeline only; not product embedder). Do not mix with Reddy 0.849 / US 0.500 / ES n=62 first-stage. Plan: [22-multilingual-ecommerce-search.md](22-multilingual-ecommerce-search.md).
+
+**Production Italian + US n=74 arms (2026-08-19).** Isolation held: `searchbench*` only; `_SEARCH_DDL` stays `to_tsvector('english'`; locomo/beam/`demorecsys` never get `search_tsv_alt`. **Decision:** production Italian is a dedicated `searchbench*` brain + `SEARCH_FTS_BRAINS` + `SEARCH_FTS_REGCONFIG=italian` + optional `extras={"locale":"it"}`. Alt BM25 matches **any** query lexeme (OR); english `search_tsv` keeps `plainto_tsquery` AND. Inflect skip-ingest `search-italian-smoke-inflect-fts-or`: BM25 nonempty for `bollitori` / `divani` / `caffettiere`. Do not quote nDCG as Italian quality.
+
+US n=74 skip-ingest, passages, k=50, `rerank=none` (control nDCG@10 **0.500** / Recall@10 **0.379** / Recall@50 **0.834** / p50 ~60 ms):
+
+- Frozen-head cascade is in `hybrid.frozen_head_merge` and is wired only when extra plugin (or `SEARCH_LITERAL_FILL`) lists are present. Default omitted `channels=["passages"]` stays hybrid RRF. Replay of stored passages+BGE+ColBERT lists still Recall@50 **0.889** / nDCG@10 **0.500**. Live n=74 sidecar not run: ColBERT/SPLADE plugin indexes are empty for `searchbenchesci74`. Honest default Recall@50 stays **0.834**.
+- `fusion=cc` alpha `{0.3,0.5,0.7}`: best nDCG@10 **0.493** (alpha 0.7) vs win **0.520**. **Null.** Leave RRF.
+- Title-token literal residual (`SEARCH_LITERAL_FILL`, frozen-head): Recall@10 **0.379** held; Recall@50 **0.655** (below 0.834). **Null.** Flag stays false.
+- Head LTR (`./search.sh ltr-head`, run `search-esci-74-ltr-head-k50`): query-grouped 5-fold CV reorders stored `search-esci-74-passages-k50` `hit_ids` (same 50; Recall@50 **0.834** held by construction). nDCG@10 **0.515**, Recall@10 **0.384** vs win **0.520** / **0.397**. Overlap-only diagnostic nDCG@10 **0.476** / Recall@10 **0.339**. **Null.** Not live. Do not train-on-all-74 as the quality number.
+- Head LTR + 4-class `ce_gain` feature (`search-esci-74-ltr-cefeat-k50`): same CV, pair policy `other_query_neg`. nDCG@10 **0.524** (gate ≥ 0.520), Recall@10 **0.383** (hold ≥ 0.379), Recall@50 **0.834**. **Harness win on predeclared nDCG/hold.** Not live. Human review before `search.py`. Do not mix with Reddy 0.857 or CE-as-sole-ranker (those hurt Recall@10).
+- Head LTR + DeBERTa-v3-base `ce_gain` (`search-esci-74-ltr-deberta-k50`): same lists/CV/`other_query_neg`; checkpoint `data/models/esci-deberta-v3-base-4class`. nDCG@10 **0.542** (must-beat this round: > **0.524**), Recall@10 **0.387** (≥ 0.379), Recall@50 **0.834**. **Harness win vs MiniLM blend.** Horizon nDCG@10 ≥ **0.70** missed (oracle on these 50 is 0.876). Gated LightGBM lambdarank (`search-esci-74-ltr-deberta-lgbm-k50`): nDCG@10 **0.516**, Recall@10 **0.379** (unrounded 0.3788, below hold), `ce_gain` gain-importance largest. **Null vs RankNet DeBERTa.** Not live. Do not mix with Reddy 0.857, pool nDCG@20 0.710, or DeBERTa-as-sole-sorter (0.510 / 0.363).
+- I04 train-on-matched-hybrid (`search-esci-74-ltr-deberta-train200`): RankNet+DeBERTa fit on 170 US train queries' hybrid k=50 from new brain `searchbenchesciltr2` (JSONL cap 200q/4k docs, doc budget bound to **170** queries / **4000** docs; 0 qid overlap with the 74). Applied to frozen `search-esci-74-passages-k50`. nDCG@10 **0.544** (> **0.542**), Recall@10 **0.391** (≥ 0.379), Recall@50 **0.834**. **Harness win vs I01 CV.** Horizon 0.70 missed. Train-set first-stage diagnostic (different catalog): nDCG@10 **0.387** / Recall@50 **0.744**. Not live. Do not wipe `searchbenchesci74`.
+- I-LGBM-APPLY (`search-esci-74-ltr-deberta-lgbm-train200`): LightGBM lambdarank on the same 170 hybrid lists / DeBERTa `ce_gain` / `other_query_neg`, applied to frozen 74. Frozen hypers (`n_estimators=100`, `max_depth=3`, `lr=0.05`). nDCG@10 **0.533**, Recall@10 **0.389**, Recall@50 **0.834**. **Null vs RankNet apply 0.544.** Horizon 0.70 missed. Stop-for-human bar 0.58 not reached. Not live.
+- I-CE-HYB (`search-esci-74-ltr-deberta-hybrid170`): continue pool DeBERTa 4-class on 8500 hybrid-k50 rows from the 170 lists (`data/models/esci-deberta-v3-base-4class-hybrid170`; did not clobber the pool checkpoint). RankNet apply with new `ce_gain`. nDCG@10 **0.530**, Recall@10 **0.385**, Recall@50 **0.834**. **Null vs C6 0.544.** Horizon 0.70 missed. Not live.
+
+Product default that survived: hybrid BM25+dense, passages, `rerank=none`, extras/facets as shipped, Italian via gated FTS+OR. Harness nDCG@10 **0.544** is not a live default. Do not mix with Reddy 0.857 or CE-on-pool 0.710.
+
+## WANDS first-stage control (2026-08-20)
+
+**Decision:** frozen `data/search_wands.jsonl` (caps 80q/2000d/40 cand; **actual 66 queries / 2000 docs**), brain `searchbenchwands`, passages, k=50, `rerank=none`, skip_enrichment, no graph. Run `search-wands-passages-k50`. Gains Exact=1 / Partial=0.5 / Irrelevant=0. Recall is binary Exact+Partial. Linear DCG. After first eval, `--skip-ingest` only. `download --name wands` refuses overwrite.
+
+**Located evidence:** nDCG@10 **0.823**, Recall@10 **0.269**, Recall@50 **0.837**, MRR **0.925**, p50 **87 ms**. Gold median 40; Recall@10 ceiling **0.365**. Do **not** mix with ESCI 0.500 / C6 0.544, Reddy 0.857, or Chen et al. (DOI located; nDCG **not located**). Details: [23-search-wands-quality.md](23-search-wands-quality.md).
+
+## Opt-in catalog graph (2026-08-20)
+
+**Decision:** architecture-opt-in on new brain `searchbenchwandsgraph` only. Product mapper is [`src/core/search/catalog_graph.py`](../../src/core/search/catalog_graph.py) (harness re-exports it). Communities match hubs then intersect item neighbors when ≥2 hub kinds match, else union; hybrid-rank product nodes inside `graph_channels`. Live omitted `channels` stays `["passages"]`. Node BM25 is `SEARCH_ENABLED` english `kg_nodes.search_tsv`. Do not edit `search.py` fusion this round.
+
+**Isolation:** reuse frozen `data/search_wands.jsonl`; never wipe `searchbenchwands` / `searchbenchesci74`. Ledger runs `search-wandsgraph-communities-k50` / `search-wandsgraph-passages-k50` are labeled **architecture-demo**. Do not quote them as beating 0.823 / 0.500. Details: [24-search-catalog-graph.md](24-search-catalog-graph.md).
+
+## Query-gated personalize (2026-08-21)
+
+**Decision:** optional `SearchRequestBody.target` reranks retrieved hits after extras AND. Omit `target` → anonymous ranking (this protocol’s default). Do not call `/retrieve/recommend`. Do not score HitRate as nDCG. If ledgered: `architecture-demo` only. Details: [25-personalized-search.md](25-personalized-search.md), [ADR-008](../decisions/008-query-gated-search-personalization.md).
diff --git a/docs/research/19-search-esci-quality.md b/docs/research/19-search-esci-quality.md
new file mode 100644
index 0000000..e28e28e
--- /dev/null
+++ b/docs/research/19-search-esci-quality.md
@@ -0,0 +1,1166 @@
+# 19 — ESCI slice search quality: retrieval, fusion, and eval
+
+Workstream: `/retrieve/search` ranking quality on `searchbench*` only. Ledger: `benchmarks.search` only. This document is a **proposal and experiment plan**, not a published finding.
+
+**Focal question:** Which interventions could raise nDCG@10, Recall@10, and/or MRR on BrainAPI search for the ESCI product-search slice, and which of those are actually testable given the current eval?
+
+Session purpose: decide the next search-quality experiments. Audience: product owner. Decision owner: the user. Horizon: days.
+
+Every claim is labeled **idea**, **assumption**, **prediction**, **located evidence**, or **decision**.
+
+---
+
+## What this workstream does
+
+Search quality here is labeled product ranking, not LoCoMo judge accuracy and not `/retrieve/recommend`.
+
+### Write path (this slice)
+
+1. Catalog JSONL docs are ingested as chunks with a `DOCID <doc_id>` marker (`benchmarks/search/catalog.py:234-256`, `dataset.py:9-12`). Default ingest is `skip_enrichment=true` (chunk + embed only; `evaluate.py:45-46`, `cli.py:46-47`).
+2. Optional `--ingest-graph` writes deterministic HAS triples (`evaluate.py:315-320`). Entity uuid equals catalog `doc_id` (`mapping.py:39-40`). Hubs are `hub:{kind}:{value}` (`mapping.py:43-44`). Class/brand/color/locale/feature values become CLASS / ATTR / TYPE nodes (`mapping.py:180-251`). If no fields exist, a TYPE hub from `dataset` is used (`mapping.py:241-250`).
+3. **Located evidence:** this slice’s 200 docs have brand/color/locale in the JSONL and **0/200 `class` fields**. CLASS hubs from mapping are therefore mostly absent; ATTR/TYPE hubs still exist.
+
+### Read path
+
+1. Default omitted `channels` is `["passages"]` (`src/services/api/constants/requests.py:478`; `hooks.py:81-89` treats core names as non-plugins).
+2. Passages: dense ANN + BM25, fused with RRF or convex combination (`search.py:176-177, 257-265`; `hybrid.py:109-135`; RRF at `fact_filter.py:67-78`).
+3. Graph channels (`graph_channels.py:18-28`):
+ - `entities`: name CONTAINS via `graph.search_entities` plus optional node ANN (`graph_channels.py:280-392`). No default `node_labels` filter unless the request sets one (`search.py:111`).
+ - `events`: same as entities with `EVENT` labels plus recency (`graph_channels.py:395-433`). Empty-ok on catalog-only brains.
+ - `communities`: search TYPE/CLASS/TOPIC hubs (default `SEARCH_COMMUNITY_LABELS`, `config.py:701-709`; `graph_channels.py:30, 564-615`), then expand neighbors with degree-IDF. Member hits can be ATTR hubs because skip labels are only the community labels (`graph_channels.py:588-614`).
+4. Each populated graph channel is appended as its own `extra_id_lists` entry and RRF-fused with passage id lists (`search.py:238-265`). IDs are mixed chunk ids and node uuids. A unit test locks this mixing in (`tests/test_search_graph.py:409-416`).
+5. Optional `rerank=plugin:<name>` reranks at most `RERANK_MAX_K = 10` on `mode=default` (`hooks.py`; `search.py`). Opt-in `mode=catalog` retrieves `min(200, max(k, 50))` and reranks at most `CATALOG_RERANK_MAX_K = 50`, then cuts to request `k`. `rerank=none` on both ESCI arms unless stated.
+
+### Eval path
+
+1. After ingest (or `--skip-ingest` reuse), chunks are listed and mapped by DOCID substring (`evaluate.py:332-335`, `dataset.py:61-76`).
+2. Gold is doc_id space: `gold_hit_ids` unions `gold_doc_ids` / positive `gold_grades` and maps `gold_chunk_ids` through `chunk_to_doc` (`evaluate.py:201-222`).
+3. Hits are canonicalized chunk to doc and deduped; unknown ids (including `hub:*`) stay as themselves (`evaluate.py:149-161, 361-362`).
+4. Metrics: Recall@{5,10,20}, graded nDCG@10, MRR (`metrics.py:7-67`). nDCG uses `gold_grades` when present (`metrics.py:42-56`).
+5. Ledger upserts `benchmarks.search` and now records `channels` (`report.py:61, 157`).
+
+### Located evidence — two arms on the same brain
+
+Same dataset `benchmarks/data/search_esci_slice.jsonl` (200 docs, 11 queries, all product columns), brain `searchbenchesci20`, fusion `rrf`, rerank `none`, skip_enrichment true. Graph arm ingested triples; passages arm used `--skip-ingest` (chunks reused; `eval.json` `ingest.reused=true`).
+
+| Arm | Channels | nDCG@10 | Recall@10 | Recall@20 | MRR | p50 retrieve |
+| --- | --- | --- | --- | --- | --- | --- |
+| Graph | passages+entities+communities | 0.6815 | 0.4657 | 0.5818 | 0.8485 | 108 ms |
+| Passages | passages | 0.7579 | 0.5106 | 0.8470 | 0.8485 | 29 ms |
+
+Sources: `benchmarks/runs/search-esci-slice-allcols/{report,eval}.json`, `benchmarks/runs/search-esci-slice-passages/{report,eval}.json`.
+
+Do **not** claim graph-channel lift. Graph dropped nDCG@10 (−0.076) and Recall@10 (−0.045). MRR was identical on every query. Recall@20 dropped far more (−0.265).
+
+Per-query first-relevant rank never moved. Five graph queries put `hub:attr:*` in the fused top-10. Graph unique hit counts after canonicalize were 12–18 vs 20 for passages. Offline drop of `hub:*` from the already fused graph lists only moved nDCG 0.682 to 0.691 and Recall@10 0.466 to 0.476 (Recall@20 unchanged). **Prediction that failed:** removing visible hubs from the scored list recovers passages. Residual gap is ASIN substitution and a shorter unique product pool created at fusion time.
+
+One query (esci-18) had graph nDCG +0.057 with unchanged Recall@10 because rank-10 swapped a Substitute (gain 0.1) for an Exact (gain 1.0). That is a single-query graded accident, not a mechanism claim.
+
+Both arms score 0 on esci-72 (`$100 things that are not electronics`; 6 Exact golds never retrieved).
+
+Gold sizes: min 1 / median 20 / max 20. Gain histogram on the slice: E=1.0 ×127, S=0.1 ×37, C=0.01 ×2. Recall@10 is therefore capped near 0.5 on the typical query even if top-10 is all gold.
+
+---
+
+## Maintainer decision (2026-08-18)
+
+**Decision (located):** make graph channels useful even if passages stays the default. Headline nDCG vs passages-only is **not** the go/no-go. Isolated-channel gold retrieval is: Recall@10 of gold ASINs and presence of product uuids in hits. Follow-up evals reuse brain `searchbenchesci20` with `--skip-ingest`. Default omitted `channels` remains `["passages"]`.
+
+---
+
+## Isolated channels then ranking change (same brain, n=11)
+
+All arms: `search_esci_slice.jsonl`, `--skip-ingest`, fusion `rrf`, rerank `none`. No re-ingest. `eval.json` now stores per-hit `id`, `channel`, and canonical `doc_id` (`evaluate.py`).
+
+### Before ranking change (measurement)
+
+| Arm | Channels | nDCG@10 | Recall@10 | Recall@20 | MRR | p50 ms | Queries with gold in top-10 | `hub:*` in top-10 |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| Passages | passages | 0.758 | 0.511 | 0.847 | 0.848 | 29 | 10/11 | 0 |
+| Entities | entities | 0.532 | 0.415 | 0.709 | 0.667 | 53 | 10/11 | 6 queries (13 slots) |
+| Communities | communities | 0.000 | 0.000 | 0.000 | 0.000 | 52 | 0/11 | empty lists |
+| Passages+entities | passages,entities | 0.682 | 0.466 | 0.582 | 0.848 | 74 | 10/11 | 5 queries |
+
+Sources: `runs/search-esci-slice-{passages,entities,communities,pe}/eval.json`.
+
+**Call (located evidence):** entities-only already retrieved gold ASINs (Recall@10 0.415, 10/11 queries). ATTR hubs stole rank on 6 queries. Communities-only never fired (TYPE/CLASS/TOPIC hubs do not match these queries; this slice has 0/200 `class` fields). Passages+entities matched the original fused graph arm — communities added nothing. Peer RRF of entity lists with passages still shortened the product pool (Recall@20 0.582 vs 0.847).
+
+Offline compacting of `hub:*` from the original fused `hit_ids` (promote later ranks into @10) yields nDCG@10 0.764 / Recall@10 0.476 / Recall@20 0.582 / MRR 0.848 (`python -m search.replay_fusion`). An earlier 0.691 nDCG figure did not refill @10 from ranks 11+. Compact nDCG matching passages is **not** a claim that live fusion is fixed; Recall@20 stays 0.582.
+
+### After ranking change (same frozen corpus)
+
+Code: `graph_channels.py` defaults the entities channel to `ENTITY` nodes, ranks by name overlap, drops `hub:*` / ATTR/TYPE/CLASS members from entity lists. Communities also seed ATTR hubs (brand/color/feature), expand at most 8 hubs, and emit only item ENTITY neighbors. `search.py` will not RRF hub ids into fused lists. Default omitted channels still `["passages"]`.
+
+| Arm | Channels | nDCG@10 | Recall@10 | Recall@20 | MRR | p50 ms | Queries with gold in top-10 | `hub:*` in top-10 |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| Passages smoke | passages | 0.758 | 0.511 | 0.847 | 0.848 | 28 | 10/11 | 0 |
+| Entities | entities | 0.655 | 0.482 | 0.822 | 0.757 | 51 | 10/11 | 0 |
+| Communities | communities | 0.478 | 0.350 | 0.391 | 0.848 | 129 | 10/11 | 0 |
+| Fused | passages,entities,communities | 0.653 | 0.497 | 0.653 | 0.816 | 184 | — | 0 |
+
+Sources: `runs/search-esci-slice-{passages-smoke,entities-after,communities-after,pec-after}/eval.json`.
+
+**Useful, in numbers (go/no-go):** isolated entities Recall@10 0.415 → 0.482; zero hub ids in top-10; product uuids on all 11 queries; Recall@20 0.709 → 0.822. Isolated communities Recall@10 0 → 0.350 with gold in top-10 on 10/11 queries (esci-19 still 0). Passages-only smoke is unchanged at 0.758 / 0.511 / 0.848.
+
+**Not a quality win for fused search.** p+e+c nDCG@10 0.653 is still below passages 0.758 (and below the old fused 0.682). MRR fell from 0.848 to 0.816. Do not put entities/communities in the default mix. Do not claim ESCI-wide lift. n=11.
+
+Communities retrieve p50 129 ms is under the ~200 ms SLO; an uncapped hub expansion on this brain was 239 ms p50 and was capped to 8 hubs.
+
+### Still not useful / unmeasured
+
+- **events:** no interaction EVENT ingest on this catalog; still empty-ok.
+- **expand=neighbors:** unit-tested; no isolated ESCI arm.
+- **esci-72:** entities-only still 0 gold in top-10; passages-only still 0; communities-after had 1/6 gold in top-10 — n=1, not a negation fix.
+- **Fused graph+passages:** still crowds Recall@20 vs passages (0.653 vs 0.847).
+
+Reproduce (no re-ingest):
+
+```bash
+cd benchmarks
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest --fusion rrf \
+ --channels entities --run search-esci-slice-entities-after
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest --fusion rrf \
+ --channels communities --run search-esci-slice-communities-after
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest --fusion rrf \
+ --channels passages --run search-esci-slice-passages-smoke
+```
+
+---
+
+## Passages CE, esci-72, ranking-in-pool (2026-08-18)
+
+**Decision:** default `rerank=none` stays. Isolated graph channels stay off the default mix.
+
+### Passages MiniLM (Task 1)
+
+`runs/search-esci-slice-passages-ce` — `--channels passages --rerank plugin:cross-encoder`, `--skip-ingest`, n=11.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | MRR | p50 retrieve |
+| --- | --- | --- | --- | --- | --- |
+| Passages (control) | 0.758 | 0.511 | 0.847 | 0.848 | 28–29 ms |
+| Passages + MiniLM k≤10 | **0.738** | 0.511 | 0.847 | 0.803 | 30 ms |
+
+**Located evidence:** nDCG@10 −0.020 vs control (win rule was +0.02). Recall@20 unchanged. MRR down. p50 retrieve 30 ms passes the ~200 ms labeled gate. p95 client wall ~7.5 s is first-query model load, not retrieve. **Null result.** Do not change the product default.
+
+### esci-72 dump (Tasks 2–3)
+
+Query: `$100 things that are not electronics`. All 6 Exact golds are **in the JSONL**. None appear in passages top-20. Helper: `benchmarks/search/miss.py`.
+
+| ASIN | Title (short) | Marker |
+| --- | --- | --- |
+| B00TREI0JI | Apple iPad Air 2 (Renewed) | electronics |
+| B073R68TSH | Beam Electronics car phone holder | electronics |
+| B07NSX2ZBS | FindKey RF key/phone tracker | electronics |
+| B07QYLTT7X | Galaxy Watch Active TPU case | electronics |
+| B07XG6Y847 | AirPods silicone case | electronics |
+| B083S3ZXDF | Potaroma electric flopping fish toy | electronics (electric toy) |
+
+Hits are mostly price tags, gift cards, Pokémon lots, batteries — lexical `$100` / “tags”, not a negation failure. **Decision: qrel mismatch.** Do not rewrite `not electronics` on this slice. No eval-only intervention.
+
+### Ranking-in-pool (Tasks 4–7)
+
+Catalog query rows now store `candidate_doc_ids` / `candidate_grades` (I=0 kept off `gold_grades`). Slice JSONL grew 200 → 365 docs so CE-on-pool has texts for I-labels. Those extra docs were **not** ingested into `searchbenchesci20`.
+
+| Arm | Protocol | nDCG@10 | nDCG@20 | nDCG full | R@10 | MRR | notes |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| BrainAPI in-pool | search then keep pool ids | 0.788 | **0.811** | 0.814 | 0.535 | 0.909 | coverage 0.47; ~15.5 pool ids/query missing from brain |
+| MiniLM CE-on-pool | harness CE on all labeled pairs | 0.598 | **0.669** | 0.786 | 0.399 | 0.833 | `cross-encoder/ms-marco-MiniLM-L-6-v2`; missing_text=0 |
+
+**Cite Reddy EN CE 0.857 / KDD 0.9043 only against CE-on-pool.** That arm is n=11, zero-shot MS MARCO, not ESCI-finetuned, not the 4,477-query public test. Do not average these rows with slice first-stage 0.758.
+
+Reproduce:
+
+```bash
+cd benchmarks
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest \
+ --channels passages --rerank plugin:cross-encoder \
+ --run search-esci-slice-passages-ce
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest \
+ --channels passages --rank-pool \
+ --run search-esci-slice-passages-pool
+./search.sh --brain searchbenchesci20 rank-pool-ce \
+ --dataset data/search_esci_slice.jsonl \
+ --run search-esci-slice-ce-pool
+```
+
+`rank-pool-ce` uses the repo `.venv` (sentence-transformers). 74-query ingest is a later human checkpoint on a **new** `searchbench*` brain.
+
+---
+
+## Guarantees and where they break
+
+**Stated guarantee (this workstream):** hybrid search returns a ranked list of catalog items such that graded nDCG@10, Recall@10, and MRR measure product relevance on ESCI-style qrels, with p50 retrieve (ex-embed) under ~200 ms (`18-search-eval-protocol.md` SLO).
+
+Where that guarantee cannot be delivered today:
+
+1. **Fusion pollution (high impact on nDCG and recall).** RRF treats graph node lists as peer ranked lists (`search.py:246-265`). Incompatible ids occupy slots; more importantly, extra lists reorder and crowd out passage ASINs. **Gap**, not a documented trade-off. Tests currently require mixed ids (`test_search_graph.py:409-416`).
+2. **Community/entity channels unfiltered (high impact on recall@20).** Entities search all labels (`search.py:111`). Communities emit neighbors of TYPE/CLASS/TOPIC, including ATTR hubs (`graph_channels.py:588-614`). Those ids cannot match gold doc_ids unless they are product entity uuids.
+3. **Eval cannot attribute a hit to a channel (high impact on diagnosis).** `eval.json` stores canonicalized `hit_ids` only (`evaluate.py:364-372`). We cannot yet falsify “graph never fired useful hits” vs “graph fired gold ASINs that RRF buried” without a new dump or isolated arms.
+4. **n=11, no pre-registered control until this ablation (critical for inference).** Means have no interval. One query (esci-67) has a single gold; one (esci-72) is a total miss. Deliberate slice size for ingest cost; **not** a full-ESCI claim.
+5. **MRR is saturated and insensitive here (important).** Mean 0.848 is “first Exact usually at rank 1.” Interventions that only clean ranks 2–10 will not move MRR. Eval artifact, not a ranking win.
+6. **skip_enrichment + generic HAS triples (medium).** Graph is attribute hubs, not LLM entity resolution. Deliberate cost trade-off. Do not interpret this graph as GraphRAG.
+7. **Title/class BM25 confound (medium, weaker than assumed on this file).** Session prior: class words in titles can inflate BM25. Located evidence: 0/200 docs have a `class` field; titles still contain type words (envelopes, pads, fence). Allcols text also includes brand and bullets. Do not attribute passages quality to graph CLASS hubs.
+8. **k=10 vs gold size ~20 (important).** Recall@10 cannot exceed ~0.5–0.67 on most queries. Recall@20 is the more sensitive recall readout on this slice (located evidence: passages 0.847 vs graph 0.582).
+
+---
+
+## Open questions for the maintainer
+
+1. Is the decision criterion “beat passages-only 0.758 / 0.511 / 0.848 on this slice,” or “make graph channels eventually useful even if passages stay the default”? **Answered:** useful isolated channels; passages stays default. Fused nDCG is not go/no-go.
+2. May all follow-up arms reuse `searchbenchesci20` with `--skip-ingest` (no re-ingest, no wipe)? **Answered:** yes.
+3. When isolated `entities` / `communities` arms show no gold ASINs, should those channels stay off the default search path (`channels=["passages"]`)? **Answered:** they stay off the default path even after they retrieve gold ASINs.
+4. Is p50 retrieve ~200 ms ex-embed still binding for `rerank=plugin:cross-encoder` on this brain?
+5. Should headline decisions use nDCG@10, or also treat Recall@20 as primary given gold size ≈20?
+6. Is `benchmarks/data/search_esci.jsonl` (title-only, 2000 docs) in scope only as a later confirmation condition, not a substitute for this allcols slice?
+7. Should `entities` default to `node_labels=["ENTITY"]` in product search, or remain unfiltered so ATTR/TYPE name matches can enter RRF?
+
+---
+
+## Independent ideas (generated before literature)
+
+Origin: AI-assisted, stage `independent` unless marked `post-check`.
+
+| ID | Statement | Cluster | Predicted metric move vs passages | Disconfirming observation |
+| --- | --- | --- | --- | --- |
+| I01 | Treat passages-only as the quality baseline; do not add graph fusion until pollution is diagnosed | eval / fusion | nDCG/recall stay; graph arm must not be the default | Isolated graph arms beat passages on nDCG and recall without hub ids |
+| I02 | Dump per-hit `channel` + run isolated `entities`, `communities`, `passages+entities` on this brain | eval | none (measurement) | Isolated arms empty and fused lists contain no graph-origin gold ASINs |
+| I03 | Never emit `hub:*` into `extra_id_lists`; only product ENTITY uuids | fusion pollution | +nDCG, +recall@10/@20; MRR flat | Offline/online filter leaves Recall@20 much less than 0.85 (already partly located) |
+| I04 | Force `node_labels=["ENTITY"]` on the entities channel | ingest/graph | +recall if entities retrieve gold ASINs; else null | Entities-only with ENTITY labels has Recall@10 ≈ 0 |
+| I05 | Downweight or gate graph lists in RRF (not peer k=60 lists) | fusion | +nDCG if entities have some gold | Weighted RRF still loses Recall@20 to passages |
+| I06 | Cross-encoder rerank on passages-only head (k≤10) | precision-at-top | +nDCG; MRR maybe; recall@10 unchanged unless head misses gold | nDCG flat after rerank; or p50 retrieve >200 ms |
+| I07 | Handle exclusionary queries (`without` / `not`) | recall | +nDCG/recall on esci-72 and similar | Rewrites still miss the 6 gold ASINs |
+| I08 | Fielded / title-weighted BM25 | precision | +nDCG if body tokens pollute | Title-only BM25 ≤ allcols passages |
+| I09 | SPLADE / ColBERT first-stage plugins | recall | +recall@10/@20 vs passages | Plugin index + latency cost, no metric lift |
+| I10 | Enlarge n + paired tests; keep this slice as a smoke | eval artifact | none (inference) | Larger ESCI still shows the same fusion drop |
+| I11 | Passages+entities without communities | fusion | smaller drop than full graph if communities are the pollutant | p+e equals or worse than p+e+c |
+| I12 | Raise retrieve `k` then cut @10 | recall (gaming) | +recall@10 by stuffing | nDCG@10 falls; SLO missed — do not use as a quality claim |
+| I13 | LLM enrichment / richer triples | ingest | unknown | Cost/latency; skip_enrichment was deliberate |
+| I14 | Use graph as expansion after passages, not as RRF peers | fusion | recover passages nDCG; optional recall if expansion adds gold | Expansion still injects non-gold ASINs into @10 |
+| I31 | Offline RRF replay from stored per-channel lists (`post-check`) | fusion | distinguishes pollution vs never-fired | Cannot replay until I02 dumps raw lists |
+
+### Criteria (declared before scoring)
+
+Scale 1–5, higher better. Weights (decision aid only; **decision stays null**):
+
+| Criterion | Weight | 1 | 3 | 5 |
+| --- | --- | --- | --- | --- |
+| information_gain | 3 | No mechanism split | Splits one pair of stories | Splits pollution vs never-fired vs passages headroom |
+| relevance_metrics | 3 | Will not move the three metrics vs 0.758/0.511/0.848 | Moves one metric on this slice | Directly targets a measured gap |
+| originality | 1 | Standard IR | Known method, new here | No nearby literature located |
+| feasibility | 2 | New models / re-ingest | 3–5 files + one eval | Eval-only on `searchbenchesci20` |
+| rigor | 2 | Easy to game | Needs care | Clean ablation, hard to HARKing |
+| value_if_null | 2 | Null is noise | Null slightly useful | Null decides the next fork |
+| latency_slo | 2 | Likely blow ~200 ms p50 retrieve | Borderline | No extra retrieve cost |
+
+Weighted score = 100 × Σ p_j (x−1)/4, p_j = w_j / Σ w. Not a winner picker.
+
+| ID | IG | Rel | Orig | Feas | Rig | Null | Lat | Score | Uncertainty |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| I02 | 5 | 4 | 3 | 5 | 5 | 5 | 5 | 92 | Low |
+| I01 | 5 | 4 | 2 | 5 | 5 | 5 | 5 | 90 | Low |
+| I04 | 4 | 4 | 2 | 5 | 4 | 5 | 5 | 82 | Med: depends on search_entities quality |
+| I03 | 4 | 5 | 3 | 4 | 4 | 4 | 5 | 82 | Low–med: post-hoc hub drop already weak |
+| I10 | 4 | 3 | 2 | 4 | 5 | 5 | 5 | 77 | Low |
+| I14 | 4 | 4 | 3 | 3 | 4 | 4 | 3 | 67 | Med |
+| I05 | 3 | 4 | 2 | 3 | 3 | 4 | 5 | 63 | Med |
+| I07 | 4 | 4 | 3 | 2 | 3 | 4 | 3 | 60 | High transfer |
+| I08 | 3 | 3 | 2 | 3 | 3 | 3 | 5 | 55 | Med |
+| I06 | 3 | 4 | 2 | 3 | 3 | 3 | 2 | 50 | Lat veto possible |
+| I09 | 3 | 4 | 2 | 2 | 3 | 3 | 2 | 47 | Fashionable here |
+| I12 | 2 | 3 | 1 | 5 | 1 | 2 | 4 | 42 | Rigor veto |
+| I13 | 2 | 2 | 2 | 2 | 2 | 3 | 4 | 40 | Out of first wave |
+
+---
+
+## Adversarial review (shortlist)
+
+**I02 isolated arms.** Falsifier: all isolated graph arms return gold ASINs in top-10 and fused p+e+c still loses — then the bug is specifically three-list RRF, not empty channels. Alternative: `--skip-ingest` passages arm hit a warmer cache (esci-67 passages retrieve 266 ms vs graph 108 ms). Residual: n=11.
+
+**I03 hub filter.** Falsifier: already located — dropping `hub:*` after canonicalize does not recover Recall@20. Alternative: hubs in the scored list are a symptom; the cause is extra RRF lists changing passage ranks. Revise I03 to “do not append non-product lists,” not “strip hubs after the fact.”
+
+**I04 ENTITY-only.** Falsifier: entities-only Recall@10 ≈ 0. Alternative: `search_entities` CONTAINS on long titles already returns products; label filter changes little. Selection: we only observed fused lists, not raw entity lists.
+
+**I06 rerank.** Falsifier: nDCG flat because first-stage already puts Exact at rank 1 (MRR 0.85). Alternative: any nDCG lift is E vs S reordering on n=11. Latency: CE on 10 docs can blow the retrieve budget; call it out. Metric gaming: rerank cannot raise Recall@10 if gold is outside the head.
+
+**I07 negation.** Falsifier: gold products for esci-72 are not matchable in this 200-doc pool. Alternative: conceptual query, not syntactic negation.
+
+**I12 raise k.** Reject as a quality intervention. It games Recall@k without fixing ranking.
+
+---
+
+## Frontier techniques
+
+Search date: **2026-08-18**. Tools: arXiv MCP `search_papers` / `get_abstract`. Queries: `"Shopping Queries Dataset" OR ESCI "product search"`; `"reciprocal rank fusion"`; `ti:"reciprocal rank fusion"`; `"knowledge graph" retrieval product search OR "entity-oriented search"`; `ti:"passage re-ranking" BERT OR ti:SPLADE OR ti:ColBERT`; `query negation retrieval OR exclusionary queries`; plus direct ids `2206.06588`, `1805.07591`, `2404.16130`, `1901.04085`, `2004.12832`, `2107.05720`. Limits: one session, cs.IR-heavy; absence from search is not novelty. Cormack et al. SIGIR 2009 RRF was not located on arXiv in this pass (`search-incomplete` for that specific paper).
+
+### Reciprocal rank fusion of heterogeneous lists
+
+- **Mechanism:** Unweighted RRF sums 1/(k+rank) across lists (`fact_filter.py:67-78`, k=60). BrainAPI appends each graph channel as an extra list (`search.py:246-265`).
+- **arXiv:** Canonical 2009 paper not retrieved here. **2503.20698** (Samuel et al., MMMORRF) reports weighted, modality-aware RRF and +37% nDCG@20 vs single-modality on video benchmarks — different domain.
+- **Cost:** Unweighted extra lists are cheap; they still change the top-k composition.
+- **Fit:** We already fuse incompatible id spaces. Weighted/gated RRF is an adapt if isolated arms show useful product ids. Peer unweighted fusion of hubs is challenge-located by our ablation (graph worse than passages).
+- **Verdict:** **adapt** gating/weighting after I02; do not add more unweighted lists.
+- **Literature status:** `mixed` (method exists; our setting violated the “compatible lists” assumption).
+
+### ESCI / Shopping Queries Dataset
+
+- **Mechanism:** Difficult Amazon queries; E/S/C/I labels; Task 1 is ranking with gains typically E=1, S=0.1, C=0.01, I=0 (`2206.06588`, Reddy et al.).
+- **Reported gain:** Paper is a dataset plus baselines, not a BrainAPI method. About 130k queries / 2.6M judgments, multilingual.
+- **Cost:** Full ESCI ingest is large; this slice is 11 queries / 200 docs by design (`catalog.py` caps).
+- **Fit:** Our grades match Task 1. Our n does not. Title-only 2000-doc file is a different condition.
+- **Verdict:** **adopt** as the label protocol; **reject** treating this slice as a full-ESCI result.
+- **Literature status:** `support-located` for metric definitions; `challenge-located` for generalizing n=11.
+
+### Entity-oriented / KG-in-the-ranker (not hub-in-RRF)
+
+- **Mechanism:** EDRM (`1805.07591`, Liu et al.) represents queries/docs with words and entity annotations; KG semantics live in embeddings; ranking is interaction-based, end-to-end.
+- **Reported gain:** Commercial search log; paper claims better generalization. Not ESCI.
+- **Cost:** Needs entity linking plus a learned ranker, not `hub:attr` uuids in RRF.
+- **Fit:** Current mapping emits generic HAS hubs (`mapping.py:180-251`) and fuses raw node ids. That is not EDRM.
+- **Verdict:** **reject** as a next experiment; **adapt** later only if entities-only retrieves gold products and we add a ranker that consumes annotations.
+- **Literature status:** `mixed`.
+
+### GraphRAG community summaries
+
+- **Mechanism:** `2404.16130` (Edge et al.): LLM-extracted entity graph, community summaries, map-reduce answers for global questions.
+- **Reported gain:** Comprehensiveness/diversity on query-focused summarization, ~1M-token corpora — not nDCG of product ASINs.
+- **Cost:** LLM index plus per-query summary tokens. Violates skip_enrichment and the search retrieve SLO.
+- **Fit:** BrainAPI `communities` are typed catalog hubs, explicitly not Leiden / `kg_topic_sessions` (`18-search-eval-protocol.md`).
+- **Verdict:** **reject**. Assumption BrainAPI violates: search is item ranking, not corpus QFS; skip_enrichment forbids the LLM graph.
+- **Literature status:** `challenge-located` (wrong task).
+
+### Cross-encoder passage rerank
+
+- **Mechanism:** `1901.04085` (Nogueira and Cho): BERT scores query–passage pairs and reorders a first-stage head. Protocol already has `plugins/search-rerank` via `rerank=plugin:cross-encoder` (`18-search-eval-protocol.md`).
+- **Reported gain:** +27% relative MRR@10 on MS MARCO passage vs prior SOTA (2019). Not ESCI; our MRR is already 0.85.
+- **Cost:** Second-stage model on ≤10 candidates (`hooks.py:7`). Likely material retrieve latency. Call out if p50 approaches 200 ms.
+- **Fit:** Best tested on **passages-only**. Reranking a polluted graph list first is the wrong order.
+- **Verdict:** **adapt** as a passages-only arm after I02, with an explicit latency gate.
+- **Literature status:** `support-located` for method; `mixed` for expected MRR lift here.
+
+### Learned sparse / late interaction
+
+- **Mechanism:** SPLADE (`2107.05720`, Formal et al.) learns sparse term weights for first-stage ranking. ColBERT (`2004.12832`, Khattab and Zaharia) late-interacts token vectors (MaxSim). Both exist as optional plugins in the protocol.
+- **Reported gain:** Competitive with dense/sparse SOTA on MS MARCO-class passage tasks. Extra index and FLOPs.
+- **Cost:** Own index (`POST /search-splade/index` / ColBERT index). Can blow the retrieve budget. Fashionable relative to this repo’s measured gap.
+- **Fit:** Worth a later passages-only bake-off, not a graph-fusion fix.
+- **Verdict:** **reject** for the next few days; **adapt** after passages-only headroom is measured on a larger n.
+- **Literature status:** `support-located` as IR methods; low expected gain per implementation cost here.
+
+### Exclusionary / negation queries
+
+- **Mechanism:** ExcluIR (`2404.17288`, Zhang et al.) shows first-stage retrievers fail on “what I do not want” queries; a dedicated train set helps but stays below humans; generative retrieval is claimed stronger.
+- **Reported gain:** Architecture-wide struggle on 3,452 annotated exclusionary queries; training helps, gap remains.
+- **Cost:** New query rewrite or trained retriever. Latency depends on implementation.
+- **Fit:** Several slice queries are `without` / `not` (esci-12/13/18/19/2/34/42/60/67/72). Only esci-72 is a total miss. Do not treat the whole slice as an ExcluIR replica.
+- **Verdict:** **adapt** as a targeted error analysis plus optional rewrite **after** fusion diagnosis. Status: `support-located` for the esci-72 failure mode.
+
+### Instruction-following retrievers / metadata-as-text
+
+- **Mechanism:** InF-IR (`2505.21439`) trains instruction-aware embeddings. Metadata-aware RAG (`2601.11863`) finds prefixing metadata in the chunk helps.
+- **Fit:** Allcols ingest already prefixes Title/Brand/bullets (`catalog.py:234-256`). InF-style models are a new index.
+- **Verdict:** **reject** for this sprint (`mixed` / not needed to explain graph vs passages). Reopen after I02.
+
+### Post-check ideation (after literature)
+
+- I31 (offline RRF replay) added. Weighted RRF (I05) upgraded from “fashionable” to “adapt if I02 shows useful entity ASINs” because **2503.20698** is support-located for *weighted* fusion, not for unweighted extra lists.
+- GraphRAG and EDRM were not added as winners. Absence of a “graph always helps product nDCG” paper in this search is not a gap we should fill with more channels.
+
+---
+
+## Critique of the current eval as a study
+
+### Summary
+
+Two arms on one frozen 200-doc / 11-query ESCI-US slice, same brain, same qrels, fusion RRF, no rerank, skip_enrichment. The graph arm added entities+communities and structured HAS triples. The passages arm reused chunks (`--skip-ingest`). Headline: passages 0.758 / 0.511 / 0.848 vs graph 0.682 / 0.466 / 0.848. This is a paired within-corpus contrast, not an RCT and not full ESCI.
+
+### Strengths
+
+- Same brain and corpus; passages arm did not re-ingest (reduces write-path drift).
+- Graded nDCG uses ESCI-style gains; gold collapse to doc_id is implemented (`evaluate.py:149-161, 201-222`).
+- Latency split retrieve vs embed is recorded; graph p50 108 ms and passages p50 29 ms both sit under the ~200 ms retrieve SLO.
+- Ledger now records `channels`, so the two arms cannot be silently compared as if they were the same system.
+- Per-query lists exist in `eval.json` even though `report.json` is aggregate-only.
+
+### Concerns
+
+**Critical**
+
+- n=11. No confidence interval, no multiplicity control, no pre-registered primary metric among {nDCG, Recall@10, Recall@20, MRR}. One total-miss query and one single-gold query dominate means.
+- No channel-attributed hits. The study cannot yet say whether graph lists contained gold ASINs.
+- Graph vs passages is confounded with extra RRF lists, mixed id spaces, structured ingest (graph arm only), and different retrieve-path work. The passages ablation removes the “no control” problem but does **not** isolate a single mechanism.
+
+**Important**
+
+- Recall@10 is a poor recall target when |gold|≈20; Recall@20 tells the real crowding story (0.847 vs 0.582).
+- MRR identical on all 11 queries: the study is silent on precision-at-1 differences.
+- Title-class BM25 confound remains a plausible passages explanation; this file has no `class` field, so the confound is title/body type words, not mapped CLASS hubs.
+- skip_enrichment plus generic HAS triples means a graph loss is not a loss for “KG retrieval” in the EDRM/GraphRAG sense.
+- esci-67 passages retrieve 266 ms vs graph 108 ms: cache or load difference; do not over-read p95.
+
+**Minor**
+
+- `report.json` omits per-query rows (they live in `eval.json`).
+- Default evaluate `k=20` then metrics cut at 10 (`cli.py:321`, `evaluate.py:277-279`).
+- Mixed chunk/node ids are tested as desired behavior (`test_search_graph.py:409-416`).
+
+### Recommendations
+
+1. Keep passages-only as the baseline for quality claims.
+2. Next measurements: isolated channels + per-hit channel dump + Recall@20 in the headline table.
+3. Do not ship more graph fusion until I02/I11 falsify or confirm pollution.
+4. Enlarge n before any “ESCI improved” sentence. Keep title-only 2000-doc as a separate condition.
+5. Preregister: primary = nDCG@10 vs passages; secondary = Recall@20; MRR expected flat; latency gate p50 retrieve <200 ms.
+
+### Overall assessment
+
+The numbers **can** support: on this 11-query allcols slice, adding entities+communities via unweighted RRF **hurt** nDCG@10 and recall relative to passages-only, while MRR did not change, and retrieve got slower (108 vs 29 ms p50). They **cannot** support: graph-aware search is worse on ESCI; passages is SOTA; class-word BM25 is proven; GraphRAG failed; or any intervention is validated. Offline hub-drop **can** support that visible `hub:*` ids are not the whole gap. It **cannot** support that graph never produced a useful product id.
+
+---
+
+## Implementation plan
+
+**Architecture decisions**
+
+- Measurement and ablations before ranking-code changes.
+- Reuse `searchbenchesci20` with `--skip-ingest`. Never wipe or score `locomoconv*`, `beam*`, `demorecsys`.
+- Quality claims compare to **passages-only** 0.758 / 0.511 / 0.848, not to the graph arm.
+- Core channel names stay `passages|entities|events|communities`. Default omitted channels stay `["passages"]`.
+- No edit to `~/.cursor/plans/graph-aware_search_channels_e0ddd720.plan.md`.
+
+Passages-only on this slice/brain is **done** (located evidence above). It is Task 0, not Task 1.
+
+### Phase 1: Measurement (fail fast)
+
+### Task 1: Persist per-hit channel and raw id before canonicalize
+
+**Description:** Write each hit’s `id`, `channel`, and pre-canonical id into `eval.json` queries so later arms can replay fusion and count graph-origin gold ASINs.
+
+**Acceptance criteria:**
+
+- [x] `eval.json` `queries[i].hits` includes `id`, `channel`, and canonical `doc_id`
+- [x] Existing metrics still match a recompute from those hits
+- [x] Ledger still upserts only `benchmarks.search`
+
+**Verification:**
+
+- [ ] `cd benchmarks && .venv/bin/python -c "import json; json.load(open('runs/search-esci-slice-passages/eval.json'))"`
+- [ ] `pytest tests/test_search_graph.py -q` (no harness import of `src/` from `benchmarks/search`)
+
+**Dependencies:** None
+
+**Files likely touched:** `benchmarks/search/evaluate.py`, `benchmarks/search/report.py`
+
+**Estimated scope:** S
+
+### Task 2: Isolated channel arms on the frozen brain
+
+**Description:** Three `--skip-ingest` evaluates on `searchbenchesci20` / `search_esci_slice.jsonl`: `entities`; `communities`; `passages,entities`. Do **not** pass `--ingest-graph`. Compare to passages-only.
+
+**Acceptance criteria:**
+
+- [x] Three new `runs/search-esci-slice-*` reports with `channels` recorded
+- [x] Table of nDCG@10, Recall@10, Recall@20, MRR, p50 retrieve vs passages
+- [x] Written call: useful-hits (gold ASINs in isolated lists) vs pollution (fused p+e or p+e+c worse than passages)
+
+**Verification:**
+
+```bash
+cd benchmarks
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest \
+ --channels entities --run search-esci-slice-entities
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest \
+ --channels communities --run search-esci-slice-communities
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest \
+ --channels passages,entities --run search-esci-slice-pe
+```
+
+**Dependencies:** Task 1 preferred (channel dump); can run without it
+
+**Files likely touched:** none in core (eval only); new run dirs
+
+**Estimated scope:** S
+
+### Task 3: Offline fusion-replay / hub-drop notebook in-repo
+
+**Description:** Script that reads two `eval.json` files (or Task 1 dumps) and recomputes metrics after (a) dropping `hub:*`, (b) intersecting with product ASINs, (c) if raw lists exist, RRF replay without graph lists. Check in under `benchmarks/search/` or a research helper, not repo root.
+
+**Acceptance criteria:**
+
+- [x] Reproduces compact hub-drop Recall@10 0.476 / MRR 0.848 from `search-esci-slice-allcols` (nDCG 0.764 when ranks 11+ refill @10; earlier 0.691 did not)
+- [x] Prints per-query deltas vs passages
+- [x] Exits 0 on the two existing eval files
+
+**Verification:**
+
+```bash
+cd benchmarks && .venv/bin/python -m search.replay_fusion \
+ --graph runs/search-esci-slice-allcols/eval.json \
+ --passages runs/search-esci-slice-passages/eval.json
+```
+
+**Dependencies:** None for (a)(b); Task 1 for (c)
+
+**Files likely touched:** `benchmarks/search/replay_fusion.py` (new), maybe `tests` under `benchmarks/` if present
+
+**Estimated scope:** S
+
+### Checkpoint: After Tasks 1–3
+
+- [x] We can say, in one sentence, whether entities-only retrieved gold ASINs
+- [x] We can say whether communities-only retrieved gold ASINs
+- [x] We can say whether passages+entities is closer to passages or to p+e+c
+- [x] Human reviews before any fusion code change
+- [x] If both isolated graph arms have Recall@10 = 0 and no gold ASINs: **stop graph fusion work**; go to Phase 3 (passages headroom)
+
+Entities-only Recall@10 was 0.415 **before** the ranking change (gold ASINs present; hubs stealing slots). Communities-only was 0. Passages+entities equaled p+e+c. Ranking/filter work proceeded; default channels were not changed.
+
+### Phase 2: Fusion hygiene (only if Task 2 shows useful product ids)
+
+### Task 4: Fuse only product ENTITY uuids
+
+**Description:** Before `extra_id_lists.append`, keep ids that are product entities (uuid equals a catalog doc_id / label ENTITY). Do not append ATTR/TYPE/CLASS hub ids. Default omitted channels remain passages.
+
+**Acceptance criteria:**
+
+- [x] Unit test: hub id does not appear in fused output when passages+entities+communities run on a fixture
+- [x] Mixed chunk+ENTITY uuid still allowed (existing test can stay for ENTITY ids)
+- [x] `--skip-ingest` p+e+c eval: Recall@10 moved toward passages (0.466 → 0.497); nDCG@10 did not (0.682 → 0.653). Default path unchanged.
+
+**Verification:**
+
+- [ ] `pytest tests/test_search_graph.py -q`
+- [ ] Repeat Task 2 p+e+c command after the change; compare to passages 0.758 / 0.511 / 0.847
+
+**Dependencies:** Task 2 (go/no-go), Task 1
+
+**Files likely touched:** `src/services/api/controllers/search.py`, `src/core/search/graph_channels.py`, `tests/test_search_graph.py`
+
+**Estimated scope:** M
+
+### Task 5: Optional ENTITY label default on entities channel
+
+**Description:** If Task 2 entities-only is dominated by ATTR hubs, pass or default `node_labels=["ENTITY"]` for catalog search without adding recsys field names to the API.
+
+**Acceptance criteria:**
+
+- [ ] Entities-only eval with `--node-labels ENTITY` recorded
+- [ ] If Recall@10 still ≈ 0, document I04 as null and do not change product default
+
+**Verification:** same `./search.sh --brain searchbenchesci20 evaluate --skip-ingest --channels entities --node-labels ENTITY --run search-esci-slice-entities-label`
+
+**Dependencies:** Task 2
+
+**Files likely touched:** none, or `search.py` only if default changes
+
+**Estimated scope:** S
+
+### Checkpoint: After Tasks 4–5
+
+- [ ] Graph fusion is either gated to product ids or explicitly left off the default path
+- [ ] No claim of graph lift unless nDCG and Recall@20 both beat passages on this slice
+- [ ] p50 retrieve still <200 ms
+
+### Phase 3: Passages headroom (higher leverage than more graph)
+
+### Task 6: Passages-only cross-encoder arm (latency-gated)
+
+**Description:** `--channels passages --rerank plugin:cross-encoder` on the frozen brain. Primary: nDCG@10 vs 0.758. Hard gate: p50 retrieve. If plugin missing, expect HTTP 400, not a fake ranking miss (`hooks.py:103-113`).
+
+**Acceptance criteria:**
+
+- [x] Report written even if status failed (plugin absent)
+- [x] If ok: nDCG@10, MRR, Recall@10, p50/p95 retrieve vs passages
+- [x] No graph channels on this arm
+
+**Verification:**
+
+```bash
+cd benchmarks
+./search.sh --brain searchbenchesci20 evaluate \
+ --dataset data/search_esci_slice.jsonl --skip-ingest \
+ --channels passages --rerank plugin:cross-encoder \
+ --run search-esci-slice-passages-ce
+```
+
+**Dependencies:** Checkpoint after Phase 1; open question 4
+
+**Files likely touched:** none if plugin already wired
+
+**Estimated scope:** S
+
+### Task 7: Exclusionary miss analysis (esci-72)
+
+**Description:** For esci-72 (and optionally other `without`/`not` queries), dump gold titles vs top-20 passages. Decide whether the miss is pool coverage, lexical negation, or conceptual intent. No production rewrite until the dump exists.
+
+**Acceptance criteria:**
+
+- [x] One page or eval artifact listing the 6 gold ASINs and why they missed
+- [x] Binary decision: rewrite experiment vs “not testable on this pool”
+
+**Verification:** python over `eval.json` + `search_esci_slice.jsonl` (no new ingest)
+
+**Dependencies:** None
+
+**Files likely touched:** optional helper under `benchmarks/search/`
+
+**Estimated scope:** S
+
+### Task 8: Protocol — Recall@20 in report + paired note
+
+**Description:** Surface Recall@20 in `report.json` / printed table; add a one-line paired-delta note (graph minus passages) in `report.md`. Do not invent p-values on n=11.
+
+**Acceptance criteria:**
+
+- [ ] `report.json` includes `recall@20`
+- [ ] Markdown table shows channels
+
+**Verification:** `./search.sh report --run search-esci-slice-passages`
+
+**Dependencies:** None
+
+**Files likely touched:** `benchmarks/search/report.py`
+
+**Estimated scope:** S
+
+### Checkpoint: Complete
+
+- [ ] Passages-only remains the cited baseline
+- [ ] I02 falsification path executed
+- [ ] No ranking change shipped without Task 2 go/no-go
+- [ ] Human approved before any larger ESCI ingest
+
+---
+
+## Risks
+
+| Risk | Impact | Detection |
+| --- | --- | --- |
+| HARKing a graph win from esci-18 nDCG | High | Require Recall@20 and ≥2 queries |
+| Treating n=11 as ESCI | High | Dataset label + n_queries in every sentence |
+| Re-ingest drift on searchbenchesci20 | High | `--skip-ingest` only; never wipe |
+| Ledger pollution of other suites | High | Upsert `benchmarks.search` only |
+| CE / SPLADE blow 200 ms SLO | Med | Record p50 retrieve; abort claim if ≥200 ms |
+| Metric gaming via larger k | Med | Reject I12 as quality |
+| Cache confound on latency | Low | esci-67 already odd; compare medians |
+| Changing default channels in core | High | Default stays `["passages"]`; test `test_search_graph.py:108` |
+
+---
+
+## Decision log (proposal only)
+
+- Candidates considered: I01–I14, I31, plus rejected GraphRAG/EDRM-now/I12/I13.
+- Criteria/weights: table above, set before scoring.
+- Literature and review date: 2026-08-18.
+- **Decision: useful isolated graph channels; passages remains default.** Accountable human: isolated Recall@10 / gold-ASIN presence is go/no-go; fused nDCG is not.
+- Revisit trigger: larger-n ESCI, or an events/expand arm.
+
+---
+
+## Graph fusion hygiene (2026-08-18) — G01–G08
+
+Workstream: optional graph channels in `/retrieve/search` without shrinking unique-product recall or nDCG vs passages-only. Brain `searchbenchesci20`, dataset `benchmarks/data/search_esci_slice.jsonl`, n=11, `--skip-ingest`, `--fusion rrf`, `--rerank none`. Default omitted `channels` stayed `["passages"]`. No production `RERANK_MAX_K` change. Ledger: `benchmarks.search` only.
+
+Predeclared primary: Recall@20 then nDCG@10. Live fusion change only if an arm is not worse than passages on both. Null → G08 (graph as side channel).
+
+### Task 1 — per-channel lists and unique-doc counts (located evidence)
+
+`eval.json` now stores `dense_ids`, `bm25_ids`, `entity_ids`, `community_ids`, `n_unique_docs_raw`, `n_unique_docs_canonical`, and `gold_grades`. Search responses expose `channel_lists` when the API from this checkout is running. Ranking of hits is unchanged.
+
+| Arm | Run | unique-doc@20 raw | unique-doc@20 canonical |
+| --- | --- | --- | --- |
+| Passages | `search-esci-slice-passages-gate` | 20.0 | **20.0** |
+| Fused p+e+c | `search-esci-slice-pec-lists` | 20.0 | **15.3** |
+
+**Located evidence:** passages-only unique-doc@20 stays 20 on every query. Fused p+e+c still drops unique products (mean 15.3) because chunk UUIDs and catalog `doc_id`s occupy separate RRF slots (P1).
+
+### Tasks 2–3 — offline replay (located evidence)
+
+Replay: `python -m search.replay_fusion --mode all` on the Task 1 dumps (`pec-lists` + `passages-gate`), stitching isolated `entities-after` / `communities-after` when a list is missing. Collapse maps every id to `doc_id` **before** RRF. Gated arms used the same dumps; expansion N=10 and graph weights `{0.1, 0.25, 0.5}` were predeclared.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | MRR | unique@20 |
+| --- | --- | --- | --- | --- | --- |
+| Passages | 0.758 | 0.511 | 0.847 | 0.848 | 20.00 |
+| G01 collapse-rrf | 0.723 | 0.498 | 0.839 | 0.847 | 20.00 |
+| G04 weighted-0.1 | 0.740 | 0.511 | 0.847 | 0.852 | 20.00 |
+| G04 weighted-0.25 | 0.725 | 0.498 | 0.847 | 0.852 | 20.00 |
+| G04 weighted-0.5 | 0.726 | 0.498 | 0.852 | 0.852 | 20.00 |
+| G02 expansion N=10 | 0.758 | 0.511 | 0.650 | 0.848 | 20.00 |
+| G05 confirmation | 0.704 | 0.502 | 0.847 | 0.795 | 20.00 |
+
+Same protocol on the earlier isolated-channel dumps (no live `channel_lists`) gave collapse R@20 0.827 / nDCG 0.686 and expansion R@20 0.815 / nDCG 0.758. Winner rule was identical.
+
+**P1 survived as the unique-doc / Recall@20 hole.** Collapse restores unique@20 to 20 and Recall@20 to 0.839 (not ≪ 0.847). **P2/P3 survived as the nDCG hole.** Collapse and every weighted/confirmation mix stay below passages nDCG 0.758. Expansion keeps nDCG but loses Recall@20 (novel graph ASINs displace gold in ranks 11–20).
+
+No arm is not-worse than passages on **both** nDCG@10 and Recall@20.
+
+### Checkpoint — G08 (decision)
+
+**Decision:** G08. Keep graph off the ranking mix; expose it as a side channel. Do not implement G01/G02/G04/G05 in `fuse_passage_lists` / live RRF. Default omitted channels remain `["passages"]`.
+
+This is not a claim that graph lists contain no gold ASINs (isolated entities Recall@10 0.482 still holds). It is a claim that **unweighted or gated fusion does not beat passages on this n=11 slice**.
+
+### Task 5 — live passages gate (located evidence)
+
+`runs/search-esci-slice-passages-gate` — `--channels passages`, `--skip-ingest`.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | MRR | p50 retrieve |
+| --- | --- | --- | --- | --- | --- |
+| Passages (frozen control) | 0.758 | 0.511 | 0.847 | 0.848 | 28–29 ms |
+| Passages gate (this run) | 0.758 | 0.511 | 0.847 | 0.848 | 26 ms |
+
+Within noise of the frozen control. p50 retrieve 26 ms passes the 200 ms gate. Fused p+e+c `pec-lists` is unchanged at 0.653 / 0.497 / 0.653, p50 160 ms (still under 200 ms, still worse than passages). Do not treat 0.857 as a number from this workstream (that figure is CE-on-pool only).
+
+### Task 6 — `expand=neighbors` (located evidence)
+
+`runs/search-esci-slice-entities-neighbors` — `--channels entities --expand neighbors`.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | MRR | unique@20 | p50 retrieve |
+| --- | --- | --- | --- | --- | --- | --- |
+| Entities-only (prior) | 0.655 | 0.482 | 0.822 | 0.757 | 20 | 51 ms |
+| Entities + neighbors | 0.655 | 0.482 | 0.822 | 0.757 | 20 | **223 ms** |
+
+Recall@10 matches entities-only 0.482. `neighbor_ids` were empty on all 11 queries (adjacent items already in the entity seed set, or dropped by the item filter). p50 223 ms **fails** the 200 ms gate. Stop. Do not fuse neighbors into default search.
+
+### What this does not support
+
+- Product claim of “graph precision” or ESCI-wide lift.
+- Changing default search to graph.
+- Shipping ID-collapse, weighted RRF, expansion-after, or confirmation as a ranking change.
+- n=11 as a published ESCI result.
+
+---
+
+## CE-on-pool / first-stage n=74 (2026-08-18, separate from fusion)
+
+Fusion files were not edited here. Protocol: CE-on-pool vs Reddy **0.857**; first-stage vs **0.495** nDCG@10. n=74 ≠ ~4477.
+
+**Located evidence**
+
+- Live `--skip-ingest --channels passages` on `searchbenchesci74`: nDCG@10 **0.495**, nDCG@20 0.549, R@10 0.377, R@20 0.592, MRR 0.758, p50 62 ms (matches the prior first-stage row).
+- Local fielded BM25 / RM3 / SPLADE on the same 2043-doc JSONL: best nDCG@10 **0.465** (SPLADE) / **0.458** (title-boost BM25). All **below** 0.495. First-stage win rule (≥0.515) **failed**.
+- 4-class MiniLM-L-12 CE-on-pool (40k US Task-1 train, test qids held out, catalog text, weighted ESCI gains): nDCG@20 **0.678**, full-list nDCG **0.747**. Exceeds our FT **0.669**. Does **not** beat Reddy **0.857**.
+
+Do not mix these rows with slice n=11 0.758. Do not change product default rerank.
+
+---
+
+## Deeper k, then harness CE on retrieved hits (2026-08-18)
+
+**Decision:** Goal 1 is more gold in the retrieved list (Recall@k, pool coverage). Goal 2 is nDCG of that same list. Graph stays off the ranking mix (G08). Production default `k` and `RERANK_MAX_K = 10` were not changed. Do **not** cite Reddy **0.857** on these arms (that figure is CE-on-pool only).
+
+**Control (located evidence):** `searchbenchesci74`, `search_esci_74.jsonl`, `--skip-ingest --channels passages --fusion rrf --k 20`: nDCG@10 **0.495**, nDCG@20 0.549, Recall@10 0.377, Recall@20 0.592 (`search-esci-74-passages-control`). With `--rank-pool`: nDCG@20 0.579, **pool_coverage 0.517**, `missing_from_brain` 0 (`search-esci-74-passages-pool`).
+
+**Predeclared Phase 1 win:** pool_coverage **≥ 0.57** or Recall@50 **≥ 0.65**. nDCG@10 may stay flat. **Phase 2** only after that win: nDCG@10 **≥ 0.515** on the same retrieve-k list, Recall@50 not down vs Phase 1.
+
+### Phase 1 — deeper first-stage retrieve (located evidence)
+
+Same brain, `--skip-ingest`, passages only, `rerank=none`. Mean `n_hits` at k=50 is 50; unique-doc@50 is 50. k=100 unique-doc@100 is 100. p50 retrieve 58–60 ms (under 200 ms).
+
+| Arm | k | Recall@10 | Recall@20 | Recall@50 | Recall@100 | nDCG@10 | nDCG@20 | p50 ms |
+| --- | --- | --- | --- | --- | --- | --- | --- | --- |
+| Control | 20 | 0.377 | 0.592 | — | — | **0.495** | 0.549 | 62 |
+| `search-esci-74-passages-k50` | 50 | 0.379 | 0.590 | **0.834** | — | 0.500 | 0.552 | 60 |
+| `search-esci-74-passages-k100` | 100 | 0.379 | 0.591 | 0.831 | **0.906** | 0.500 | 0.552 | 58 |
+
+`--rank-pool` (coverage of the labeled candidate list, not Reddy Task 1):
+
+| Arm | k | pool_coverage | missing_from_brain | Recall@20 | Recall@50 | nDCG@10 | nDCG@20 |
+| --- | --- | --- | --- | --- | --- | --- | --- |
+| `search-esci-74-passages-pool` | 20 | **0.517** | 0 | 0.592 | — | 0.546 | 0.579 |
+| `search-esci-74-passages-k50-pool` | 50 | **0.778** | 0 | 0.669 | 0.832 | 0.567 | 0.624 |
+| `search-esci-74-passages-k100-pool` | 100 | **0.846** | 0 | 0.696 | 0.906 | 0.574 | 0.635 |
+
+**Phase 1 call:** win. Recall@50 **0.834 ≥ 0.65**; pool_coverage **0.778 ≥ 0.57**. Extra golds sit in ranks 21–50 (Recall@20 is flat vs k=20). nDCG@10 is unchanged (~0.500). k=100 does not raise Recall@50 vs k=50; Recall@100 is 0.906. Do not mix with CE-on-pool **0.695** or Reddy **0.857**.
+
+### Phase 2 — harness CE over stored `hit_ids` (located evidence)
+
+Sibling of `rank_pool.py`: `benchmarks/search/rerank_retrieved.py` reorders the k=50 retrieved list. It does not call production rerank and does not score the full labeled pool. Default API `rerank=none` and `RERANK_MAX_K = 10` unchanged.
+
+| Arm | model | nDCG@10 | nDCG@20 | Recall@10 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- | --- | --- |
+| k=50 first-stage | none | **0.500** | 0.552 | 0.379 | **0.834** | 0.765 |
+| `search-esci-74-passages-k50-ce` | MS MARCO MiniLM-L-6 | 0.448 | 0.515 | 0.331 | 0.834 | 0.720 |
+| `search-esci-74-passages-k50-ce-l12` | 4-class MiniLM-L-12 e2 | 0.467 | 0.508 | 0.341 | 0.834 | 0.706 |
+
+**Phase 2 call:** null. nDCG@10 did not reach **0.515** (needed +0.02 vs the 0.495 k=20 control). Both CEs **hurt** nDCG@10 vs first-stage k=50. Recall@50 held at 0.834 (same hit set). These rows are first-stage retrieve-then-rerank, not CE-on-pool 0.695 / Reddy 0.857.
+
+**Decision:** keep production search at default k / `RERANK_MAX_K=10`. Deeper retrieve finds more gold; zero-shot and local 4-class CE over that retrieved list do not promote it into the nDCG@10 head.
+
+---
+
+## Gold in small and large k (2026-08-18)
+
+**Focal question:** raise gold in both Recall@10 and Recall@50 on the same 74 queries. Graph off. Production `k` / `RERANK_MAX_K` unchanged. Do **not** cite Reddy **0.857** or CE-on-pool **0.695**.
+
+**Predeclared dual win:** Recall@10 **≥ 0.397** and Recall@50 **≥ 0.854** (both +0.02 vs k=20 / k=50 passages). Secondary nDCG@10 ≥ 0.515.
+
+### Task 1 — miss taxonomy (located evidence)
+
+`miss-strata` on `search-esci-74-passages-k50` / `search_esci_74.jsonl`. Reconciles **74** queries / **1270** golds.
+
+| Stratum | Queries | Gold items |
+| --- | --- | --- |
+| head-ok (gold in top-10) | 67 | 424 in top-10 |
+| rank-too-low (gold in 11–50 only) | 2 | 728 in ranks 11–50 |
+| total-miss | 5 | 118 missed |
+
+Total-miss qids: `esci-113`, `esci-177`, `esci-267`, `esci-393`, `esci-72`. Dual-k gap is mostly **rank-too-low golds** (728), not the five total misses. Macro Recall@50 0.834 vs micro miss 118/1270 (9%).
+
+### Task 2 — query-side on pathological total-miss only (located evidence)
+
+Rewrote **only** `esci-113` (HTML unescape) and `esci-267` (strip leading `- *`). Did **not** rewrite `esci-72` (prior qrel mismatch). SQL-ish `esci-177` left unchanged. Live `--skip-ingest` `search-esci-74-passages-k50-qrewrite`:
+
+| Arm | Recall@10 | Recall@50 | nDCG@10 |
+| --- | --- | --- | --- |
+| k=50 passages | 0.379 | 0.834 | 0.500 |
+| qrewrite (2 qids) | 0.379 | 0.834 | 0.500 |
+
+Per-qid Recall@10/@50 for `esci-113` and `esci-267` stayed **0**. Stratum-only null. Global dual win not moved.
+
+### Task 3 — local ANCE dual encoder (located evidence)
+
+Harness MiniLM-L-6 dual encoder, US Task-1 train, test qids held out, hard negatives from stored k=50 non-gold `hit_ids` (bank 1387), 6557 triples, 1 epoch. Encode 2043 JSONL docs locally. Not BrainAPI embeddings. Run `search-esci-74-dense-ance-k50`.
+
+| Arm | Recall@10 | Recall@50 | nDCG@10 |
+| --- | --- | --- | --- |
+| Passages k=20 control | **0.377** | — | **0.495** |
+| Passages k=50 | 0.379 | **0.834** | 0.500 |
+| Local ANCE MiniLM | 0.302 | 0.750 | 0.396 |
+
+**Dual-k call:** null. Both recalls moved **down**. Do not fuse this retriever into passages. Do not raise `RERANK_MAX_K`. ColBERT and BGE sidecars were run next (below); both also null.
+
+**Decision:** BrainAPI hybrid passages remain the first-stage control. Matching interventions tried here (query rewrite on 2 qids; MiniLM ANCE on the 2043-doc JSONL) did not put more gold in both small and large k.
+
+---
+
+## Retrieved-neg CE, then ColBERT, then BGE (2026-08-18)
+
+**Focal question:** (1) promote Exact already in hybrid ranks 11–50 via a 4-class CE trained on BM25 retrieved negatives; (2–3) try two first-stage sidecars that are not MiniLM dual-encoder. Graph off. Frozen `searchbenchesci74` / `--skip-ingest`. `RERANK_MAX_K` stayed 10. Do **not** mix these rows with CE-on-pool **0.695** or Reddy **0.857**. n=74, no p-values. All three arms were run even though arm 1 was null.
+
+**Predeclared wins:** arm 1 nDCG@10 **≥ 0.515** and Recall@50 not below **0.834**. Arms 2–3 Recall@10 **≥ 0.397** **and** Recall@50 **≥ 0.854**. Secondary nDCG@10 ≥ 0.515.
+
+### Arm 1 — retrieved-negative 4-class CE (located evidence)
+
+Train lists: US Task-1 small-train, test qids held out (`n_holdout_qids=148`), BM25 top-50 over the train product catalog, unlabeled hits = class **I**. 6000 queries, 300000 pairs, all lists had ≥1 I. Checkpoint `benchmarks/data/models/esci-minilm-l12-retrieved` (`source=retrieved-bm25`, 80k pairs, 1 epoch, unweighted). Eval: harness `rerank-retrieved` on stored k=50 `hit_ids` (`search-esci-74-passages-k50-ce-retrieved`). Not production rerank.
+
+| Arm | nDCG@10 | Recall@10 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- |
+| k=50 passages | **0.500** | 0.379 | **0.834** | 0.765 |
+| retrieved-neg 4-class L-12 | 0.465 | 0.351 | 0.834 | 0.712 |
+
+**Arm 1 call:** null. nDCG@10 **0.465** did not reach **0.515** and is below first-stage 0.500 (same direction as pool-trained / MS MARCO CE on these hits). Recall@50 held at **0.834** (same hit set).
+
+### Arm 2 — ColBERT MaxSim sidecar (located evidence)
+
+In-process `plugins/search-colbert` (`colbert-ir/colbertv2.0`) over the 2043 JSONL docs. Brain id `harness-local-colbert`. Did **not** `POST /search-colbert/index` onto `searchbenchesci74`. Run `search-esci-74-colbert-k50`. Channels `harness-colbert`. Not fused with passages.
+
+| Arm | Recall@10 | Recall@50 | nDCG@10 | p50 encode+retrieve |
+| --- | --- | --- | --- | --- |
+| Passages k=20 / k=50 | **0.377** / 0.379 | **0.834** | **0.495** / 0.500 | ~60 ms retrieve |
+| ColBERT MaxSim | 0.311 | 0.714 | 0.434 | 10063 ms |
+
+**Arm 2 call:** dual-k null. Recall@10 **0.311** < 0.397; Recall@50 **0.714** < 0.854. p50 ~10 s is **labeled**, not a 200 ms product claim, not a stop.
+
+### Arm 3 — zero-shot BGE-base (located evidence)
+
+`BAAI/bge-base-en-v1.5` via `local-dense` on the same 2043 JSONL. **Not MiniLM.** No `query:` prefixes. Run `search-esci-74-bge-base-k50`. No new production brain.
+
+| Arm | Recall@10 | Recall@50 | nDCG@10 |
+| --- | --- | --- | --- |
+| Passages k=20 / k=50 | **0.377** / 0.379 | **0.834** | **0.495** / 0.500 |
+| MiniLM ANCE (prior, not this arm) | 0.302 | 0.750 | 0.396 |
+| BGE-base (this arm) | 0.322 | 0.799 | 0.441 |
+
+**Arm 3 call:** dual-k null. Recall@10 **0.322** < 0.397; Recall@50 **0.799** < 0.854. Better than MiniLM ANCE 0.302 / 0.750, still below hybrid passages. p50 encode+retrieve 408 ms.
+
+**Three-arm call:** no arm beat the passages control on its predeclared gate. Keep default omitted `channels` = passages. Do not raise `RERANK_MAX_K`. Do not fuse ColBERT or BGE into passages. Do not wipe `searchbenchesci74`.
+
+---
+
+## Exhaustive catalog rank then two-stage (2026-08-19)
+
+**Focal question:** (A) does scoring **every** JSONL doc with the unweighted 4-class L-12 e2 checkpoint beat hybrid k=50; (B) do complementary sidecar lists raise dual-k recall without diluting Recall@10; (C) ship opt-in `mode=catalog` on `/retrieve/search` without changing the default path. Graph off. Frozen `searchbenchesci74` / `--skip-ingest`. Default `RERANK_MAX_K` stayed **10**. Do **not** mix these rows with CE-on-pool **0.695** or Reddy **0.857**. n=74, no p-values. B and C ran even though A was null.
+
+**Predeclared gates:** A nDCG@10 **≥ 0.515** **and** Recall@10 **≥ 0.397**. B Recall@10 **≥ 0.397** **and** Recall@50 **≥ 0.854**, and Recall@10 not below **0.379**. C (same k=50 hits, deeper rerank): A’s nDCG@10 gate with Recall@50 held at **0.834**.
+
+### A — exhaustive 4-class on 2043 docs (located evidence)
+
+Harness `rank-corpus` scored all 2043 JSONL passages per query with `benchmarks/data/models/esci-minilm-l12-4class-nowt-e2` (74 × 2043 ≈ 151k pairs). Protocol `exhaustive-catalog`. Did **not** POST onto `searchbenchesci74`. Run `search-esci-74-exhaustive-ce`. Channels `exhaustive-4class`. Cut @ 5,10,20,50.
+
+| Arm | nDCG@10 | Recall@10 | Recall@50 | p50 |
+| --- | --- | --- | --- | --- |
+| k=50 passages | **0.500** | **0.379** | **0.834** | 60 ms retrieve |
+| Exhaustive 4-class L-12 e2 | 0.416 | 0.266 | 0.603 | 94 s labeled |
+
+**A call:** null. nDCG@10 **0.416** < **0.515** and below hybrid 0.500. Recall@10 **0.266** < **0.397**. p50 is **labeled**, not a 200 ms product claim. MiniLM 4-class over the full file does not beat hybrid RRF on this slice.
+
+### B1 — U01 unique-gold overlap (located evidence)
+
+`list-overlap` on stored `hit_ids` (`search-esci-74-passages-k50` vs `search-esci-74-bge-base-k50` / `search-esci-74-colbert-k50`). Gold ASINs in sidecar top-50 absent from passages k=50:
+
+| Sidecar | unique gold hits | queries with unique golds |
+| --- | --- | --- |
+| BGE-base k=50 | **40** | 19 / 74 |
+| ColBERT k=50 | **21** | 13 / 74 |
+
+Unique golds were nonzero, so B2 included ColBERT.
+
+### B2 — harness RRF union (located evidence)
+
+RRF of passages k=50 + BGE k=50 + ColBERT k=50. Channels `harness-union`. Not fused into live graph. Run `search-esci-74-union-bge-k50`.
+
+| Arm | nDCG@10 | Recall@10 | Recall@50 |
+| --- | --- | --- | --- |
+| k=50 passages | **0.500** | **0.379** | **0.834** |
+| Union passages+BGE+ColBERT | 0.479 | 0.363 | 0.827 |
+
+**B call:** null. Recall@10 **0.363** is below control **0.379** (graph-style dilution). Dual-k **≥0.397 and ≥0.854** failed. Extra lists steal top-10 slots; unique golds in the tail do not raise Recall@50 here (0.827 vs 0.834). Do not fuse these lists into production ranking.
+
+### C — live `mode=catalog` (located evidence)
+
+Shipped: `SearchRequestBody.mode` default `"default"` (retrieve `k`, rerank `min(10, len)`). `mode=catalog` retrieves `k_ret = min(200, max(k, 50))`, reranks `min(len, CATALOG_RERANK_MAX_K=50)`, cuts to request `k`. Live `--skip-ingest` on frozen `searchbenchesci74`, `SEARCH_RERANK_MODEL` = nowt-e2 (same checkpoint as A, not retrieved-neg). Run `search-esci-74-catalog-ce-k50`. Request `k=50` so first-stage depth matches the k=50 control; this tests **plugin rerank of 50** (not 10).
+
+| Arm | nDCG@10 | Recall@10 | Recall@50 | p50 retrieve | p50 client wall |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages (`rerank=none`) | **0.500** | **0.379** | **0.834** | 60 ms | — |
+| `mode=catalog` + plugin CE | 0.467 | 0.341 | **0.834** | 59 ms | 877 ms |
+
+**C call:** architecture **shipped**; quality gate **null**. Recall@50 held at **0.834** (same 50-hit retrieve). nDCG@10 **0.467** < **0.515** and below first-stage 0.500 — same direction as harness CE-on-retrieved 4-class L-12 e2 (0.467). Catalog+CE is **not** the ADR-007 200 ms default. Omitted `mode` and `RERANK_MAX_K=10` stay the product path. Do not wipe `searchbenchesci74`.
+
+**A/B/C call:** exhaustive MiniLM cannot beat hybrid on this 2043-doc slice; list union dilutes Recall@10; catalog mode is the two-stage hook. Keep default omitted `channels` = passages. Do not cite Reddy **0.857** against these rows.
+
+---
+
+## Frozen-head cascade (2026-08-19)
+
+**Focal question:** can sidecar unique golds raise Recall@50 if hybrid top-10 is frozen? Graph off. Harness only. Not live fusion. `RERANK_MAX_K` unchanged. Do **not** mix with CE-on-pool **0.695** or Reddy **0.857**. n=74, no p-values.
+
+**Predeclared gate:** Recall@10 **≥ 0.379** (no dilution) **and** Recall@50 **≥ 0.854**.
+
+**Located evidence.** `cascade-lists`: passages ranks 1–10 copied; unique golds from BGE+ColBERT top-50 that were absent from passages k=50 inserted at 11–50; remaining passages tail golds kept before non-golds. Run `search-esci-74-cascade-tail-k50`. Channels `harness-cascade`. Protocol `frozen-head-cascade`. Injected **51** unique gold ASINs (BGE 40 ∪ ColBERT 21).
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | Recall@50 | nDCG@20 |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages | **0.500** | **0.379** | 0.590 | 0.834 | 0.552 |
+| RRF union (prior, not this arm) | 0.479 | 0.363 | 0.577 | 0.827 | 0.538 |
+| Frozen-head cascade | **0.500** | **0.379** | **0.750** | **0.889** | **0.619** |
+
+**Call:** win. Recall@10 **0.379** matches the passages control to floating point (head frozen). Recall@50 **0.889 ≥ 0.854**. nDCG@10 held; nDCG@20 rose because extras land in 11–20. RRF union failed by stealing the head; cascade uses the same unique golds without that. Not fused into live `/retrieve/search`. Do not raise default `RERANK_MAX_K`. Do not wipe `searchbenchesci74`.
+
+---
+
+## DeBERTa-v3-base 4-class on catalog k=50 (2026-08-19)
+
+**Focal question:** does replacing MiniLM with DeBERTa-v3-base (same 4-class weighted-gain recipe) promote golds already in hybrid ranks 11–50 into the top-10? Graph off. Frozen `searchbenchesci74` / no wipe. Default `RERANK_MAX_K` stayed **10**. Live `mode=catalog` only if harness gates fired. Do **not** mix first-stage rows with CE-on-pool **0.695** / **0.710** or Reddy **0.857**. n=74, no p-values.
+
+**Predeclared first-stage win:** Recall@10 **≥ 0.397** **and** nDCG@10 **≥ 0.515**, Recall@50 held at **0.834**.
+
+**Train (located).** `finetune-4class` `--base microsoft/deberta-v3-base`, US Task-1 train, test qids held out (`n_holdout_qids=148`), 80k pairs, 2 epochs, batch 8, unweighted, max_length 192, label smoothing 0.1. Checkpoint `benchmarks/data/models/esci-deberta-v3-base-4class`. Parent venv needed `sentencepiece` (installed 0.2.2). HuggingFace DeBERTa-v3 tokenizer shipped `model_max_length` as an unbounded sentinel; it was set to **192** to match train. Not v3-large. Not the www ensemble.
+
+**Pool discriminator (ranking-in-pool only; not the R@10 gate).** `rank_pool_4class` run `search-esci-74-ce-pool-deberta-base`. nDCG@20 **0.710** vs MiniLM 4-class e2 **0.695**; full-list nDCG **0.777** vs **0.765**. n=74 ≠ ~4477. Do not cite Reddy **0.857** against this row.
+
+**First-stage harness (primary).** `rerank-retrieved` on stored `search-esci-74-passages-k50` `hit_ids`. Run `search-esci-74-passages-k50-ce-deberta`. Plugin 4-logit softmax gains. No live API.
+
+| Arm | nDCG@10 | Recall@10 | Recall@50 |
+| --- | --- | --- | --- |
+| k=50 passages | **0.500** | **0.379** | **0.834** |
+| MiniLM 4-class e2 on those hits | 0.467 | 0.341 | **0.834** |
+| DeBERTa-v3-base 4-class on those hits | 0.510 | 0.363 | **0.834** |
+
+**Call:** first-stage **null**. Recall@10 **0.363 < 0.397** (also below passages 0.379). nDCG@10 **0.510 < 0.515** (above MiniLM 0.467 and slightly above passages 0.500). Recall@50 **0.834** held (same hit set). Live `mode=catalog` was **not** run: MiniLM catalog already matched harness CE-on-retrieved, and both quality gates missed. Cascade remains the Recall@50 win. Do not raise `RERANK_MAX_K`. Do not wipe `searchbenchesci74`.
+
+**Interpretation (not a finding beyond this slice).** Pool nDCG@20 moved up a little while first-stage Recall@10 moved down vs hybrid: extra backbone capacity helped ranking-in-pool more than promoting golds through the hybrid head. Compatible with the alternative that labeled pools (coverage 1.0) are easier than the k=50 retrieved list (coverage 0.778). Zhang **0.90** is still ranking-in-pool + large + ensemble, not this y-axis.
+
+---
+
+## Multilingual first-stage (2026-08-19)
+
+Not a US n=74 quality rerun. Not Reddy 0.857. Details in [22-multilingual-ecommerce-search.md](22-multilingual-ecommerce-search.md).
+
+**Located evidence.** ESCI locales are US/ES/JP (`product_locale` in Reddy HTML). No Italian in ESCI. Amazon-M2 has locale IT but rec/title-generation tasks (`2307.09688`). mMARCO has Italian as MT passages (`2108.13897`), not catalog qrels. arXiv `product search` + Italian + `cs.IR` returned 0.
+
+**Decision.** ES download must not clobber US JSONL. Italian smoke fixture is not ESCI. Product default stays passages + `rerank=none`. `searchbenchesci74` not wiped.
+
+**ES first-stage (located evidence).** Run `search-esci-es-passages-k50`, brain `searchbenchescies`, n=62, 2000 docs, k=50 passages, `rerank=none`. nDCG@10 **0.577**, Recall@10 **0.353**, Recall@50 **0.914**, p50 84 ms. Not US n=74 0.500/0.379. Not Reddy ES 0.849 ranking-in-pool.
+
+---
+
+## Production first-stage arms on US n=74 (2026-08-19)
+
+**Focal question:** can cheap first-stage changes raise Recall@50 and/or nDCG@10 on skip-ingest `searchbenchesci74` without CE-on-retrieved or graph-as-default? Graph off. Frozen brain / no wipe. `RERANK_MAX_K` stayed **10**. Do **not** mix with CE-on-pool **0.710** or Reddy **0.857**. n=74, no p-values.
+
+**Predeclared gates.** Hold: nDCG@10 **≥ 0.500**, Recall@10 **≥ 0.379**, p50 retrieve **< 200 ms** (ex-embed). Recall@50 win: **≥ 0.854** and hold head. nDCG@10 win: **≥ 0.520** and Recall not down. Null → leave hybrid RRF, `rerank=none`.
+
+### Frozen-head cascade (opt-in lists)
+
+**Located evidence.** Helper `frozen_head_merge` (hybrid top-10 frozen; unique extra ids fill 11–50). Replay unit test on stored `search-esci-74-passages-k50` + BGE + ColBERT reproduces Recall@50 **0.889**, nDCG@10 **0.500**, Recall@10 **0.379**. Live `/retrieve/search` uses this merge **only** when plugin retriever lists (or `SEARCH_LITERAL_FILL`) are requested — not default RRF of plugins into the head. Default omitted `channels=["passages"]` unchanged.
+
+**Live n=74 sidecar:** not run. `search-colbert` / `search-splade` load, but the ColBERT in-memory index for `searchbenchesci74` is empty (retrieve returns `[]`). Indexing 2043 chunks at query time was not started. **Decision:** cascade stays harness-only. Honest production Recall@50 on the default path stays **0.834**.
+
+### C1 — `fusion=cc` alpha sweep (null)
+
+Skip-ingest, passages, k=50, `rerank=none`. Request `fusion=cc` + `fusion_alpha`. p50 61–63 ms (hold).
+
+| Arm | nDCG@10 | Recall@10 | Recall@50 | p50 retrieve |
+| --- | --- | --- | --- | --- |
+| k=50 passages RRF | **0.500** | **0.379** | **0.834** | ~60 ms |
+| CC α=0.3 (`search-esci-74-cc-a03`) | 0.483 | 0.376 | 0.832 | 63 ms |
+| CC α=0.5 (`search-esci-74-cc-a05`) | 0.486 | 0.377 | 0.832 | 61 ms |
+| CC α=0.7 (`search-esci-74-cc-a07`) | 0.493 | 0.378 | 0.834 | 61 ms |
+
+**Call:** null. Best nDCG@10 **0.493 < 0.520**. Recall@10 slightly below hold. Product default stays RRF.
+
+### C2 — title-token literal residual (null)
+
+**Predeclared fusion:** frozen-head (not RRF) so the top-10 cannot be stolen. Env `SEARCH_LITERAL_FILL` default **false**. Live skip-ingest `search-esci-74-literal-fill-k50`.
+
+| Arm | nDCG@10 | Recall@10 | Recall@50 | p50 retrieve |
+| --- | --- | --- | --- | --- |
+| k=50 passages | **0.500** | **0.379** | **0.834** | ~60 ms |
+| Literal fill + frozen head | **0.500** | **0.379** | **0.655** | 49 ms (`search.retrieve` only on this run) |
+
+**Call:** null. Recall@10 held (frozen head). Recall@50 **0.655** dropped because non-gold ILIKE token hits filled 11–50 and evicted hybrid tail golds. Flag stays false. Do not treat this as a Recall@50 win.
+
+### C3 — pairwise LTR on the stored k=50 head (null)
+
+**Focal question:** can a cheap linear RankNet on features already in the hybrid list promote the 728 golds sitting in ranks 11–50 into the top-10, without changing the retrieved set? Graph off. Frozen `searchbenchesci74` / no wipe / skip-ingest. Not live `/retrieve/search`. Do **not** mix with CE-on-pool **0.710** or Reddy **0.857**. n=74, no p-values.
+
+**Protocol (located evidence).** `./search.sh ltr-head --from-run search-esci-74-passages-k50 --dataset data/search_esci_74.jsonl --run search-esci-74-ltr-head-k50`. Query-grouped 5-fold CV (OOF predictions; seed=0, 40 epochs, lr=0.05, l2=1e-3, max 400 pairs/query). Features: `rrf_inv`, `bm25_inv`, `dense_inv`, `title_overlap`, `brand_hit`, `query_in_title`. Unlabeled docs in the 50 scored as 0. Overlap-only sort stored as `overlap_only_metrics`, not the gated arm.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages RRF | **0.500** | **0.379** | 0.590 | **0.834** | 0.765 |
+| Overlap-only (diagnostic) | 0.476 | 0.339 | 0.580 | 0.834 | 0.732 |
+| LTR OOF (`search-esci-74-ltr-head-k50`) | 0.515 | 0.384 | 0.583 | **0.834** | 0.753 |
+| Predeclared win | **≥ 0.520** | **≥ 0.397** | — | hold | — |
+
+Mean CV weights (not a quality claim): `title_overlap` 3.35, `dense_inv` 2.06, `rrf_inv` 1.57, `brand_hit` 0.51, `bm25_inv` 0.02, `query_in_title` ~0.
+
+**Call (decision):** null. nDCG@10 **0.515 < 0.520**. Recall@10 **0.384 < 0.397**. Recall@50 held. Directional nDCG lift is small and is **not** promotion of the 728 (Recall@10 +0.005; Recall@20 and MRR slightly down vs RRF). Overlap-only is worse than RRF — naive title ranking is not the product. Default stays hybrid RRF, `rerank=none`. Do not wire this ranker into `search.py`.
+
+### C4 — LTR + 4-class CE as a feature (harness win; not live)
+
+**Focal question:** can blending the already-trained 4-class MiniLM score with hybrid RRF via query-grouped LTR raise nDCG@10 on the frozen k=50 list, with Recall@10 held? Graph off. Frozen `searchbenchesci74` / no wipe. Not live `/retrieve/search`. Do **not** mix with CE-on-pool **0.710**, Reddy **0.857**, or CE-as-sole-ranker (located evidence: Recall@10 0.331–0.351). n=74, no p-values.
+
+**Protocol (located evidence).** `./search.sh ltr-head --from-run search-esci-74-passages-k50 --ce-model data/models/esci-minilm-l12-4class-nowt-e2 --pair-policy other_query_neg --run search-esci-74-ltr-cefeat-k50`. Same hypers as C3 (seed=0, 40 epochs, lr=0.05, l2=1e-3, max 400 pairs/query, 5-fold OOF). Feature `ce_gain` = 1·P(E)+0.1·P(S)+0.01·P(C); lists are not sorted by CE. Pair policy: this-query graded golds vs other-query golds in the 50; unlabeled skipped. Predeclared this round: nDCG@10 ≥ **0.520**, Recall@10 ≥ **0.379**, Recall@50 held. Recall@10 ≥ 0.397 is not required to call the harness win.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages RRF | 0.500 | 0.379 | 0.590 | **0.834** | 0.765 |
+| CE-as-sole-ranker (prior, MiniLM on retrieved) | 0.448 | 0.331 | — | 0.834 | 0.720 |
+| LTR OOF no CE (`search-esci-74-ltr-head-k50`) | 0.515 | 0.384 | 0.583 | **0.834** | 0.753 |
+| LTR + `ce_gain` (`search-esci-74-ltr-cefeat-k50`) | **0.524** | 0.383 | 0.569 | **0.834** | **0.779** |
+| Predeclared this round | **≥ 0.520** | **≥ 0.379** | — | hold | — |
+
+Mean CV weights (not a quality claim): `ce_gain` 3.68, `title_overlap` 2.56, `dense_inv` 1.23, `rrf_inv` 0.66, `brand_hit` 0.45, `bm25_inv` −0.13, `query_in_title` ~0.
+
+**Call (located evidence + decision):** harness win on the predeclared nDCG/hold. nDCG@10 **0.524 ≥ 0.520**. Recall@10 **0.383 ≥ 0.379**. Recall@50 held. MRR 0.765 → 0.779. LightGBM not run (gated only if nDCG close-but-short). **Do not wire into `search.py` without human review:** live CE on 50 hits is extra latency vs the ~60 ms retrieve path; n=74 has no CI; Recall@20 fell 0.590 → 0.569. CE-alone remains a null. Default stays hybrid RRF, `rerank=none` until review.
+
+**A/B/C call:** Italian OR-FTS is a pipeline feature on an allowlisted searchbench. US n=74 live default is unchanged: hybrid BM25+dense, passages, `rerank=none`. Harness nDCG@10 0.524 is not Reddy **0.857**. Later DeBERTa RankNet blend (C5) is the current harness best at **0.542**; still not live.
+
+### C5 — DeBERTa `ce_gain` LTR blend (harness win vs MiniLM; not live)
+
+**Focal question:** on frozen [`search-esci-74-passages-k50`](../../benchmarks/runs/search-esci-74-passages-k50), does replacing MiniLM `ce_gain` with the existing DeBERTa-v3-base 4-class checkpoint raise first-stage nDCG@10 **above 0.524**, with Recall@10 ≥ **0.379** and Recall@50 **0.834** held? Graph off. Frozen `searchbenchesci74` / no wipe. Not live `/retrieve/search`. Do **not** mix with Reddy **0.857**, pool nDCG@20 **0.710**, or DeBERTa-as-sole-sorter (located: nDCG@10 **0.510**, Recall@10 **0.363**). n=74, no p-values. Horizon nDCG@10 ≥ **0.70** is not a fail of this arm.
+
+**Protocol (located evidence).** `./search.sh ltr-head --from-run search-esci-74-passages-k50 --ce-model data/models/esci-deberta-v3-base-4class --pair-policy other_query_neg --run search-esci-74-ltr-deberta-k50`. Same RankNet hypers as C3/C4 (seed=0, 40 epochs, lr=0.05, l2=1e-3, max 400 pairs/query, 5-fold OOF). Lists are not sorted by DeBERTa. Scores cached at `runs/search-esci-74-passages-k50/ce_gain_esci-deberta-v3-base-4class.json` (MiniLM cache left in place). Tokenizer already had `model_max_length=192`; `load_4class_predict` was not edited. Predeclared this round: nDCG@10 **> 0.524**, Recall@10 ≥ **0.379**, Recall@50 held.
+
+Gated I03 (beat 0.524, far from 0.70, `|ce_gain|` not ~0): LightGBM lambdarank, frozen `n_estimators=100`, `max_depth=3`, `learning_rate=0.05`, same CV/features/pair policy, run `search-esci-74-ltr-deberta-lgbm-k50`. No hypersearch.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages RRF | 0.500 | 0.379 | 0.590 | **0.834** | 0.765 |
+| DeBERTa as sole sorter (prior) | 0.510 | 0.363 | — | 0.834 | — |
+| LTR + MiniLM `ce_gain` (`search-esci-74-ltr-cefeat-k50`) | 0.524 | 0.383 | 0.569 | **0.834** | 0.779 |
+| LTR + DeBERTa `ce_gain` RankNet (`search-esci-74-ltr-deberta-k50`) | **0.542** | **0.387** | 0.586 | **0.834** | 0.774 |
+| LightGBM + DeBERTa `ce_gain` (`search-esci-74-ltr-deberta-lgbm-k50`) | 0.516 | 0.379 | 0.592 | **0.834** | 0.710 |
+| Must-beat this round | **> 0.524** | **≥ 0.379** | — | hold | — |
+| Horizon (not a fail) | **≥ 0.70** | — | — | — | — |
+| Oracle on these 50 | 0.876 | 0.548 | — | 0.834 | — |
+
+Mean RankNet CV weights (not a quality claim): `ce_gain` 3.65, `title_overlap` 2.43, `dense_inv` 1.26, `brand_hit` 0.51, `rrf_inv` 0.45, `query_in_title` 0.02, `bm25_inv` −0.08.
+
+Mean LightGBM gain importances (not a quality claim): `ce_gain` 646, `dense_inv` 561, `rrf_inv` 337, `title_overlap` 162, `brand_hit` 82, `bm25_inv` 19, `query_in_title` 0. LightGBM Recall@10 unrounded **0.3788** (quoted 0.379 in three decimals; hold is strict ≥ 0.379).
+
+**Call (located evidence + decision):** I01 **win vs MiniLM**. nDCG@10 **0.542 > 0.524**. Recall@10 **0.387 ≥ 0.379**. Recall@50 held. Horizon 0.70 missed (~0.16 short; k=50 oracle 0.876). I03 LightGBM **null**: nDCG@10 **0.516** loses to RankNet DeBERTa and to MiniLM 0.524; Recall@10 misses the hold unrounded; `ce_gain` is live, so the miss is the GBDT head, not a dead feature. **Do not wire into `search.py`:** live DeBERTa on 50 hits is extra latency vs ~60 ms retrieve; n=74 has no CI; 0.542 vs 0.524 is a point-estimate gate, not a significance test. Default stays hybrid RRF, `rerank=none` until human review. Do not replace the MiniLM cache.
+
+**I04 (this session):** both heads missed 0.70 with large `ce_gain`. Train queries on matched hybrid k=50 were run on a new `searchbench*` (C6).
+
+**A/B/C call:** US n=74 live default unchanged. Harness best after C5 was DeBERTa RankNet CV **0.542**; C6 apply-from-train is **0.544**. Still not Reddy **0.857** and still not product ranking.
+
+### C6 — LTR trained on 170 matched hybrid lists, applied to frozen n=74 (harness win vs 0.542; not live)
+
+**Focal question:** does fitting RankNet+DeBERTa `ce_gain` on **matched hybrid k=50** from extra US train queries, then applying those weights to frozen [`search-esci-74-passages-k50`](../../benchmarks/runs/search-esci-74-passages-k50), raise nDCG@10 **above 0.542**, with Recall@10 ≥ **0.379** and Recall@50 **0.834** held? Graph off. Frozen `searchbenchesci74` / no wipe. New brain only. Not live `/retrieve/search`. Do **not** mix with Reddy **0.857**, pool nDCG@20 **0.710**, or official-pool LTR. n=74, no p-values. Horizon nDCG@10 ≥ **0.70** is not a fail of 170q.
+
+**Protocol (located evidence).** Download `--split train --max-queries 200 --max-docs 4000 --out data/search_esci_ltr200.jsonl --holdout-dataset data/search_esci_74.jsonl` (did not clobber `search_esci.jsonl` / `_74`). Doc budget bound the slice to **170** queries / **4000** docs; qid overlap with the 74 is **0**. Ingest+evaluate passages k=50, `fusion=rrf`, `rerank=none`, skip_enrichment, brain **`searchbenchesciltr2`** (first `searchbenchesciltr` attempt timed out queued before the Celery worker was up; did not retry on `searchbenchesci74`). Train first-stage diagnostic (different catalog, not the gate): nDCG@10 **0.387**, Recall@10 **0.367**, Recall@50 **0.744**, p50 retrieve 76 ms (`search-esci-ltr200-passages-k50`).
+
+Then `./search.sh ltr-head --from-run search-esci-74-passages-k50 --train-from-run search-esci-ltr200-passages-k50 --train-dataset data/search_esci_ltr200.jsonl --ce-model data/models/esci-deberta-v3-base-4class --pair-policy other_query_neg --run search-esci-74-ltr-deberta-train200`. Fit **once** on 170 train lists (no CV on the 74). Reused frozen-74 DeBERTa cache; scored train lists to `runs/search-esci-ltr200-passages-k50/ce_gain_esci-deberta-v3-base-4class.json`. MiniLM cache left in place. RankNet hypers unchanged. LightGBM not re-run.
+
+**Assumption (catalog shift):** train lists come from a 4000-doc train-qrel index; test lists stay on the 2043-doc `searchbenchesci74` index. Feature space matches; IDF/dense geometry may not.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages RRF | 0.500 | 0.379 | 0.590 | **0.834** | 0.765 |
+| LTR + DeBERTa RankNet CV (`search-esci-74-ltr-deberta-k50`) | 0.542 | 0.387 | 0.586 | **0.834** | 0.774 |
+| LTR + DeBERTa applied from 170 train lists (`search-esci-74-ltr-deberta-train200`) | **0.544** | **0.391** | 0.585 | **0.834** | 0.762 |
+| Must-beat this round | **> 0.542** | **≥ 0.379** | — | hold | — |
+| Horizon (not a fail) | **≥ 0.70** | — | — | — | — |
+
+Mean applied RankNet weights (not a quality claim): `ce_gain` 4.27, `title_overlap` 3.12, `rrf_inv` 1.40, `dense_inv` 1.31, `bm25_inv` 0.49, `brand_hit` −0.23, `query_in_title` −0.22.
+
+**Call (located evidence + decision):** I04 **win vs I01** on the predeclared point-estimate gate. nDCG@10 **0.544 > 0.542**. Recall@10 **0.391 ≥ 0.379**. Recall@50 held. Horizon 0.70 missed (~0.16 short). Lift vs CV DeBERTa is **+0.002** nDCG / **+0.004** Recall@10; n=74 has no CI, so this can be noise. Train Recall@50 **0.744** shows the 170q lists are not empty of golds, but the catalog is not the 74-index. **Do not wire into `search.py`.** Default stays hybrid RRF, `rerank=none`. Next gated step if chasing 0.70 is 400q matched lists (new `searchbench*` only), not LightGBM on these 170.
+
+**A/B/C call:** US n=74 live default unchanged. Harness best is **0.544**, still not Reddy **0.857**.
+
+### C7 — LightGBM applied from 170 matched hybrid lists (null vs RankNet 0.544; not live)
+
+**Focal question:** on frozen [`search-esci-74-passages-k50`](../../benchmarks/runs/search-esci-74-passages-k50), does LightGBM lambdarank fit on the 170 train hybrid lists beat RankNet apply nDCG@10 **0.544**, with Recall@10 ≥ **0.379** and Recall@50 **0.834** held? Graph off. Frozen `searchbenchesci74` / no wipe. Not live `/retrieve/search`. Do **not** mix with Reddy **0.857** or pool nDCG@20 **0.710**. n=74, no p-values. Horizon nDCG@10 ≥ **0.70** is not a fail of this arm. Predeclared stop: nDCG@10 ≥ **0.58** would halt CE retrain.
+
+**Protocol (located evidence).** `./search.sh ltr-head --from-run search-esci-74-passages-k50 --dataset data/search_esci_74.jsonl --train-from-run search-esci-ltr200-passages-k50 --train-dataset data/search_esci_ltr200.jsonl --ce-model data/models/esci-deberta-v3-base-4class --pair-policy other_query_neg --ltr-model lightgbm --run search-esci-74-ltr-deberta-lgbm-train200`. Same DeBERTa caches as C6. Frozen LightGBM hypers. No hypersearch.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages RRF | 0.500 | 0.379 | 0.590 | **0.834** | 0.765 |
+| LightGBM 5-fold on the 74 (`search-esci-74-ltr-deberta-lgbm-k50`) | 0.516 | 0.379 | 0.592 | **0.834** | 0.710 |
+| RankNet applied from 170 (`search-esci-74-ltr-deberta-train200`) | **0.544** | **0.391** | 0.585 | **0.834** | 0.762 |
+| LightGBM applied from 170 (`search-esci-74-ltr-deberta-lgbm-train200`) | 0.533 | 0.389 | 0.591 | **0.834** | 0.752 |
+| Must-beat this round | **> 0.544** | **≥ 0.379** | — | hold | — |
+| Horizon (not a fail) | **≥ 0.70** | — | — | — | — |
+
+Mean LightGBM gain importances (not a quality claim): `ce_gain` 1684, `dense_inv` 1151, `rrf_inv` 998, `title_overlap` 273, `bm25_inv` 41, `brand_hit` 7, `query_in_title` 0.
+
+**Call (located evidence + decision):** I-LGBM-APPLY **null vs C6 RankNet**. nDCG@10 **0.533 < 0.544**. Recall@10 **0.389 ≥ 0.379**. Recall@50 held. Stop bar 0.58 not reached. Horizon 0.70 missed. Linear vs tree blender on these 170 lists is not the unused 0.33. **Do not wire into `search.py`.** Continue to hybrid-list CE as `ce_gain` (C8).
+
+**A/B/C call:** US n=74 live default unchanged. Harness best remains RankNet apply **0.544**, still not Reddy **0.857**.
+
+### C8 — hybrid-list DeBERTa as `ce_gain` (null vs C6 0.544; not live)
+
+**Focal question:** does continuing the 4-class DeBERTa on **matched hybrid k=50** lists (170 queries, unlabeled=`I`), then using the new scores as `ce_gain` in RankNet apply, beat nDCG@10 **0.544**, with Recall@10 ≥ **0.379** and Recall@50 **0.834** held? Graph off. Frozen `searchbenchesci74` / no wipe. Not live `/retrieve/search`. Do **not** mix with Reddy **0.857**, pool nDCG@20 **0.710**, or CE-as-sorter. n=74, no p-values. Horizon nDCG@10 ≥ **0.70** is not a fail of 8500 pairs.
+
+**Protocol (located evidence).** Export `./search.sh export-hybrid-lists --from-run search-esci-ltr200-passages-k50 --dataset data/search_esci_ltr200.jsonl --holdout-dataset data/search_esci_74.jsonl --out data/esci_hybrid_lists_ltr200.jsonl` (did not clobber `esci_retrieved_lists.jsonl`). 170 queries / 8500 pairs; 0 holdout-qid rows; labels E 984 / S 953 / C 117 / I 6446. Then `./search.sh finetune-4class --dataset data/search_esci_74.jsonl --from-lists data/esci_hybrid_lists_ltr200.jsonl --lists-source hybrid-k50 --base data/models/esci-deberta-v3-base-4class --out data/models/esci-deberta-v3-base-4class-hybrid170 --epochs 1 --max-length 192` (266 steps, MPS; pool checkpoint left in place; `source=hybrid-k50`). Then `./search.sh ltr-head --from-run search-esci-74-passages-k50 --train-from-run search-esci-ltr200-passages-k50 --train-dataset data/search_esci_ltr200.jsonl --ce-model data/models/esci-deberta-v3-base-4class-hybrid170 --pair-policy other_query_neg --run search-esci-74-ltr-deberta-hybrid170`. New caches only (`ce_gain_esci-deberta-v3-base-4class-hybrid170.json`). Lists not sorted by CE.
+
+**Assumption:** 8500 pairs / 170 lists may be too few to retune DeBERTa; catalog shift 4k vs 2043 remains.
+
+| Arm | nDCG@10 | Recall@10 | Recall@20 | Recall@50 | MRR |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages RRF | 0.500 | 0.379 | 0.590 | **0.834** | 0.765 |
+| RankNet + pool DeBERTa from 170 (`search-esci-74-ltr-deberta-train200`) | **0.544** | **0.391** | 0.585 | **0.834** | 0.762 |
+| LightGBM from 170 (`search-esci-74-ltr-deberta-lgbm-train200`) | 0.533 | 0.389 | 0.591 | **0.834** | 0.752 |
+| RankNet + hybrid DeBERTa (`search-esci-74-ltr-deberta-hybrid170`) | 0.530 | 0.385 | 0.588 | **0.834** | 0.765 |
+| Must-beat this round | **> 0.544** | **≥ 0.379** | — | hold | — |
+| Horizon (not a fail) | **≥ 0.70** | — | — | — | — |
+| Oracle on these 50 | 0.876 | 0.548 | — | 0.834 | — |
+
+Mean applied RankNet weights (not a quality claim): `ce_gain` 6.92, `title_overlap` 2.29, `rrf_inv` 1.14, `dense_inv` 0.84, `bm25_inv` 0.36, `brand_hit` −0.17, `query_in_title` −0.27.
+
+**Call (located evidence + decision):** I-CE-HYB **null vs C6**. nDCG@10 **0.530 < 0.544**. Recall@10 **0.385 ≥ 0.379**. Recall@50 held. Horizon 0.70 missed. One epoch of hybrid hard-negatives on 170 lists did not unlock the unused 0.33; a null here can mean too little data, not that first-stage negatives never matter. Do **not** relabel a 6k BM25-list run as this construct. **Do not wire into `search.py`.** Default stays hybrid RRF, `rerank=none`.
+
+**A/B/C call:** US n=74 live default unchanged. Harness best remains **0.544**, still not Reddy **0.857**.
+
diff --git a/docs/research/20-esci-sota-targets.md b/docs/research/20-esci-sota-targets.md
new file mode 100644
index 0000000..fb19c79
--- /dev/null
+++ b/docs/research/20-esci-sota-targets.md
@@ -0,0 +1,350 @@
+# 20 — ESCI search-quality SOTA numbers to aim for
+
+Workstream: evaluation protocol for `/retrieve/search` on `searchbench*` only. Ledger: `benchmarks.search` only. This document is **located evidence plus protocol targets**, not a ranking-code change and not a published BrainAPI finding.
+
+**Focal question:** Which published numbers are valid aim points for BrainAPI search quality, given the current 11-query / 200-doc ESCI slice?
+
+Every numeric claim below is labeled **located evidence**, **slice-internal**, or **not-checked**. Do **not** claim the slice is above or below full-ESCI SOTA.
+
+Search date: **2026-08-18**. Tools: harness read; arXiv MCP `search_papers` / `get_abstract` / `download_paper`; AIcrowd/Amazon Science web pages for leaderboard confirmation.
+
+---
+
+## What this workstream does
+
+The eval scores a ranked list from `POST /retrieve/search` against catalog qrels, not LoCoMo judge accuracy and not `/retrieve/recommend`.
+
+### Scoring path (located evidence)
+
+1. Catalog JSONL carries `gold_grades` from ESCI letters via `ESCI_GAINS = {"E": 1.0, "S": 0.1, "C": 0.01, "I": 0.0}` (`benchmarks/search/catalog.py:31`, applied at `catalog.py:298-302`). Complement **is** in the gain table. Irrelevant is dropped (`gain > 0` filter).
+2. After ingest, hits are collapsed to `doc_id` (chunk map or node uuid) and deduped (`evaluate.py:172-184`, `267-285`).
+3. nDCG@10 is **graded** when `gold_grades` is present (`metrics.py:36-56`; wired at `evaluate.py:285`). Discount is `rel / log2(rank + 1)` (`metrics.py:29-33`) — the classic Järvelin DCG, not the `2^rel − 1` variant.
+4. Recall@{5,10,20} is **binary** over the positive-gain set (Exact + Substitute + Complement), not Exact-only (`metrics.py:7-16`; gold union at `evaluate.py:224-245`).
+5. MRR is the reciprocal rank of the **first positive-gain** hit (`metrics.py:19-26`). It is not MRR@10 as a cutoff in code; the list is whatever `k` the search returned (default evaluate `k=20`, `evaluate.py:304`).
+6. Headline ledger fields are `ndcg@10`, `recall@10`, `mrr`, retrieve p50 (`report.py:65-68`).
+
+### Slice under test (located evidence)
+
+`benchmarks/data/search_esci_slice.jsonl`: 200 docs, 11 queries, US locale, all product columns, brain `searchbenchesci20`. Gold sizes: min 1 / median 20 / max 20. Gain histogram: E=127, S=37, C=2. Passages-only: nDCG@10 **0.758**, Recall@10 **0.511**, MRR **0.848**, p50 retrieve 28–29 ms (`docs/research/19-search-esci-quality.md`).
+
+Later condition, not a substitute: `benchmarks/data/search_esci.jsonl` is 2000 docs / 74 queries (title-heavy; 43/2000 docs lack Brand/Color in text; 0 `class` fields). Gold min/median/max 1/15/38, mean 17.2.
+
+Catalog builder defaults (`catalog.py:34-36, 64-71, 396-397`): locale `us`, split `test`, `small_version == 1` (Task 1 small), caps 80 queries / 2000 docs / 40 candidates per query. The JSONL is then searched as a **shared corpus**, not as a per-query candidate pool.
+
+WANDS download exists in the same harness (`catalog.py:13-32`, gains Exact=1 / Partial=0.5 / Irrelevant=0). No BEIR or TREC Product Search harness in `benchmarks/search/`.
+
+---
+
+## Guarantees and where they break
+
+**Stated guarantee:** graded nDCG@10 / Recall@10 / MRR on this slice measure product-search quality in a way that can be compared to published ESCI numbers.
+
+Where that guarantee cannot be delivered:
+
+1. **Pool size and task (critical).** Published Task 1 is *ranking-in-pool*: reorder ≤40 already-retrieved, fully labeled products per query (`2206.06588` §3.1). We retrieve from a 200-doc (or 2000-doc) shared index. Different candidate generation, different label density, different IDCG. **Gap**, not a bug in the scorer.
+2. **Cutoff k (important).** Official Task 1 is nDCG on the provided list. Reddy does not state `@10`. Jeronymo et al. report the KDD Cup metric as **nDCG@20** (`2208.06264`). We always cut at 10 (`metrics.py:121`). Same gain table, different statistic.
+3. **Recall@10 / MRR are not the Task 1 metrics (important).** Task 1 publishes nDCG. Tasks 2/3 publish micro-F1. Nobody in the primary ESCI papers reports Recall@10 or MRR as we compute them.
+4. **n=11 (critical for inference).** Means have no interval. One total miss (`esci-72`) and one single-gold query (`esci-67`) dominate averages.
+5. **Recall@10 ceiling (important).** Mean of `min(10, |gold|) / |gold|` on this slice is **0.667**. Six of eleven queries have `|gold|=20`, so Recall@10 cannot exceed 0.5 on those queries even if every top-10 hit is gold. Current 0.511 is already near that structural cap.
+6. **MRR saturation (important).** Mean 0.848 leaves little room except the total miss. Rank-2–10 work will not show up.
+
+Deliberate trade-off: slice size and `skip_enrichment` for ingest cost. Not a claim about full ESCI.
+
+---
+
+## Open questions for the maintainer
+
+1. For a confirmatory ESCI run, is the target **ranking-in-pool nDCG** on the labeled candidate list (comparable to `2206.06588` / KDD Cup), or **first-stage nDCG@10** from a shared ingested corpus?
+2. Should the headline cutoff stay nDCG@10, or switch to nDCG@20 / full-list nDCG when the goal is published-number comparability?
+3. Should Recall@10 gold stay E+S+C, or Exact-only, given Task 1’s ranking order is E then S then C then I?
+4. Is `search_esci.jsonl` (2000 docs, 74 queries) the first larger condition, or should the next run restrict scoring to each query’s original ≤40 labeled candidates?
+5. Is p50 retrieve <200 ms still binding if a confirmatory arm uses a cross-encoder the way the published 0.85–0.90 numbers do?
+
+---
+
+## Frontier techniques
+
+### ESCI / Shopping Queries Task 1 ranking-in-pool (canonical)
+
+- **Mechanism:** For each query, reorder a provided list of up to 40 products. Graded nDCG with gains E=1, S=0.1, C=0.01, I=0. Complement is in the gain table. Locales US/ES/JP. Small version (Task 1) filters easier queries; US public test is 4,477 queries, avg depth 20.3 (`2206.06588` Tables 1–4, §3.1).
+- **arXiv:** `2206.06588` (Reddy et al., 2022). Amazon Science: [dataset page](https://www.amazon.science/code-and-datasets/shopping-queries-dataset-a-large-scale-esci-benchmark-for-improving-product-search), [KDD Cup recap](https://www.amazon.science/blog/amazon-product-query-competition-draws-more-than-9-200-submissions).
+- **Reported numbers (located evidence):**
+
+| Method | Year | Metric | Split | Number | Source |
+| --- | --- | --- | --- | --- | --- |
+| Terrier BM25 (title, all locales together) | 2022 | nDCG (Task 1, no `@10` in paper) | Public test overall / EN / ES / JP | 0.563 / **0.675** / 0.697 / 0.136 | `2206.06588` Table 4 |
+| Fine-tuned MS MARCO MiniLM cross-encoder (EN) + MPNet (ES, JP); Exact→1 else 0 at train time | 2022 | nDCG | Public test overall / EN / ES / JP | **0.852** / **0.857** / 0.849 / 0.840 | `2206.06588` Table 4 |
+| Organizer “baseline” cited in the KDD recap | 2022 | nDCG | KDD private/public (unspecified cutoff in blog) | **0.8503** | [Amazon Science blog](https://www.amazon.science/blog/amazon-product-query-competition-draws-more-than-9-200-submissions) |
+| Team www (DeBERTa/XLM/RemBERT ensemble; 4-class probs → weighted gain) | 2022 | NDCG (AIcrowd Task 1) | Public / private, multilingual Task 1 | **0.9057 / 0.9043** | `2208.02958`; [AIcrowd winners](https://discourse.aicrowd.com/t/final-winners-announcement/7974) |
+| Team qinpersevere / day-day-up | 2022 | NDCG | Private Task 1 | 0.9036 / 0.9035 | AIcrowd winners table |
+| mMonoT5-3.7B (NeuralMind) | 2022 | nDCG@20 (their label for the same leaderboard) | Public / private | 0.9012 / 0.9007 | `2208.06264` Table 1 |
+| 20th place | 2022 | nDCG@20 | Private | 0.8929 | `2208.06264` Table 1 |
+| SQID reproduction of ESCI_baseline on US Task 1 test (8,956 queries, ~20 judgments/query) | 2024 | NDCG (Terrier, corrected S/C mapping) | US small test | **0.8562** | `2405.15190` Table 2 |
+| SQID random rank of the same pools | 2024 | NDCG | US small test | 0.7483 | `2405.15190` Table 2 |
+| SQID SBERT title cosine (no fine-tune) | 2024 | NDCG | US small test | 0.8292 | `2405.15190` Table 2 |
+
+- **Cost:** Cross-encoder or LLM ensembles over ~20–40 pairs per query. Not a first-stage ANN over a catalog.
+- **Fit:** Gain table matches ours. Task, n, pool, and k do not.
+- **Verdict:** **adopt** as the label protocol and as the **horizon-B ranking-in-pool target**. **reject** as a numeric bar for the 11-query / 200-doc slice.
+
+### KDD Cup 2022 as “SOTA on ESCI”
+
+- **Mechanism:** Same Task 1, private multilingual test. Top 10 private scores sit in **0.8998–0.9043** (AIcrowd). Jeronymo argues top-20 teams cluster near 0.90 nDCG@20 and that the dataset is too easy to discriminate retrievers (`2208.06264` abstract).
+- **Cost:** Heavy fine-tune + ensemble. Wu et al. (`2208.00108`) rank 6th on Task 1 but do not publish a Task 1 nDCG in the paper body (leaderboard 6th is ETS-Lab 0.9014 private; Wu’s team name is not in the top-10 table — **not-checked** which public name maps to their 6th-place claim vs AIcrowd’s ETS-Lab).
+- **Verdict:** **adopt** 0.9043 private NDCG as the published ceiling for *ranking-in-pool, multilingual Task 1*. Fashionable only if BrainAPI starts scoring labeled pools with a second-stage ranker.
+
+### Full-catalog ESCI retrieval (BLaIR)
+
+- **Mechanism:** Treat ESCI as product search over **1,367,729 items / 27,643 test queries**; report **NDCG@100** (`2403.03952` Table 3, Table 10). This is first-stage retrieval, not Task 1 reranking.
+- **Reported (located evidence):** SFR-Embedding-Mistral **0.2560**; GritLM-7B 0.2537; e5-mistral-7b-instruct 0.2437; Qwen3-Embedding-8B 0.2328; text-embedding-3-large 0.2366; gemini-embedding-001 0.2233 (`2403.03952` Table 10).
+- **Cost:** Embed the catalog; ANN. No cross-encoder over 40 candidates.
+- **Fit:** Closer to “search a brain” than Task 1, but k=100, corpus ~1.4M, and nDCG@100 ≠ our nDCG@10. Recall@10 / MRR **not reported**.
+- **Verdict:** **adapt** as the published bar if a later run is true first-stage over a large catalog. **reject** as a target for the 200-doc slice (our 0.758 lives on a different y-axis).
+
+### Papers that use ESCI but are not SOTA bars here
+
+- **ARR / SToICaL (`2601.05588`, 2026):** Custom in-context ranking, Gecko-mined lists, log-scale gains, n=310, nDCG ~95–97 and a nonstandard R@k (`|π[:k] ∩ π̂[:k]| / k`). **reject** — not ESCI Task 1 gains, not our recall.
+- **GraphRAG / EDRM / LoCoMo:** Wrong task. Out of scope.
+- **WANDS (ECIR 2022 Chen et al.):** Dataset paper identity **support-located** (2026-08-20): DOI [10.1007/978-3-030-99736-6_9](https://doi.org/10.1007/978-3-030-99736-6_9), OpenAlex `W4225565720`, Springer LNCS, `is_oa=false`, abstract index empty. arXiv title search did not return the chapter. Dataset size **located** on GitHub `wayfair/WANDS`: 480 queries / 42,994 products / 233,448 labels. Published BM25/hybrid **nDCG still not located**. Do not invent a paper nDCG. Medium notebooks are not primary. Local first-stage control is a different y-axis (≤80q / ≤2000 docs / linear Exact=1, Partial=0.5). See [23-search-wands-quality.md](23-search-wands-quality.md).
+- **TREC 2023 Product Search (`2311.07861`):** Converts SQD toward end-to-end retrieval. **No harness in this repo.** Leftover.
+- **BEIR (`2104.08663`):** No search harness here. Leftover.
+
+---
+
+## Implementation plan
+
+No ranking-code change. Protocol only. Reuse `searchbenchesci20` with `--skip-ingest`. Never wipe brains.
+
+### Phase 1 — Slice-internal aims (this week)
+
+Treat passages-only **0.758 / 0.511 / 0.848** as the control. “Beating the slice” means:
+
+| Metric | Now | Structural ceiling (slice-internal) | What a real win looks like | What is not a win |
+| --- | --- | --- | --- | --- |
+| nDCG@10 | 0.758 | Not 1.0: several queries have mixed E/S gold; `esci-72` is 0/11 of the mean | +0.02–0.05 with Recall@20 not down; or `esci-72` leaving 0 (alone +~0.09 if that query went to 1.0) | Graph fusion that drops nDCG while MRR stays flat |
+| Recall@10 | 0.511 | **0.667** mean of `min(10,\|gold\|)/\|gold\|`; **0.5** on the six `|gold|=20` queries | Move `esci-72` (0 vs ceiling 1.0), `esci-19` (0.333 vs 0.667), `esci-37` (0.667 vs 1.0) | Raising retrieve `k` then cutting @10 |
+| MRR | 0.848 | ~0.94 if only `esci-72` is 0 and is fixed to rank-1 | Almost only the total-miss query | Any rank-2–10 cleanup |
+
+**Task 1 — Publish slice ceilings next to every report**
+
+- **Acceptance:** Report or canvas shows Recall@10 ceiling 0.667 and per-query `|gold|`.
+- **Verification:** `python` over `search_esci_slice.jsonl` gold sizes (already computed 2026-08-18).
+- **Scope:** S. Files: report helper or canvas only.
+
+**Checkpoint:** Humans agree the 0.758/0.511/0.848 triple is a **control**, not a SOTA claim.
+
+### Phase 2 — Comparable confirmatory protocol
+
+Pick one of two designs. Do not mix their numbers.
+
+**Design B-rank (comparable to published 0.85–0.90):** For each query, score **only the original labeled candidates** (≤40), same E/S/C/I gains, report nDCG@20 **and** full-list nDCG, US `small_version` test, n in the thousands if ingest allows, else the 74-query 2000-doc file **restricted to each query’s gold pool**. Cite Reddy EN CE **0.857** as the organizer-style bar and www **0.9043** as the 2022 ceiling.
+
+**Design B-retrieve (comparable to BLaIR, not to 0.90):** Search a shared corpus. Report nDCG@10 plus nDCG@100 if k allows. Cite SFR-Embedding-Mistral **0.256 nDCG@100** on 1.37M items only if the corpus is actually large. The 2000-doc file is still not that condition.
+
+**Task 2 — Ranking-in-pool eval mode (if B-rank is chosen)**
+
+- **Acceptance:** Harness can restrict hits to `gold_grades` keys plus labeled irrelevants for that `qid`; metrics include nDCG@20.
+- **Verification:** One `--skip-ingest` run on a toy query with a 40-id pool; nDCG matches a manual Terrier-style recompute.
+- **Scope:** M. Files: `benchmarks/search/evaluate.py`, `metrics.py`, `report.py`.
+- **Depends:** Maintainer answer to open question 1.
+
+**Task 3 — 2000-doc condition as a separate row**
+
+- **Acceptance:** `search_esci.jsonl` run is labeled `dataset=search_esci.jsonl`, `n_queries=74`, never averaged with the 11-query slice.
+- **Verification:** `./search.sh evaluate --dataset data/search_esci.jsonl --brain searchbenchesci --skip-ingest` (or a new `searchbench*` brain — no wipe of `searchbenchesci20`).
+- **Scope:** S. Eval only.
+
+**Checkpoint:** A table with three rows that are never compared as one number: slice n=11, 2000-doc retrieve, ranking-in-pool (if implemented).
+
+### Measured n=11 ranking-in-pool smoke (2026-08-18)
+
+Not a published-number claim. Rows must not be averaged.
+
+| Row | nDCG@10 | nDCG@20 | full nDCG | n | Cite 0.857? |
+| --- | --- | --- | --- | --- | --- |
+| Slice first-stage passages | 0.758 | — | — | 11 | no |
+| BrainAPI post-filter in-pool | 0.788 | 0.811 | 0.814 | 11 | no |
+| MiniLM CE-on-pool (zero-shot MS MARCO) | 0.598 | 0.669 | 0.786 | 11 | **only this arm’s protocol**, still not the Reddy split |
+
+Harness: `--rank-pool` on evaluate; `rank-pool-ce` subcommand. Catalog persists `candidate_grades` including I=0. `searchbenchesci20` was not re-ingested.
+
+### 74-query brain `searchbenchesci74` (2026-08-18)
+
+New brain; dataset `search_esci_74.jsonl` (74 queries, 2043 pool products, mean 28.8 labeled candidates). Passages ingest onto this brain is in progress and is **not** required for CE-on-pool.
+
+Cite Reddy **0.857** only against CE-on-pool. n=74 is still not the 4,477-query public test.
+
+| Arm | nDCG@20 | full nDCG | n | vs 0.857 |
+| --- | --- | --- | --- | --- |
+| Zero-shot MS MARCO MiniLM-L-6 | 0.631 | 0.717 | 74 | below |
+| Fine-tuned 1 epoch, 30k US train pairs, Exact→1 else 0, test qids held out | 0.669 | 0.745 | 74 | below |
+
+**Located evidence:** one-epoch Exact-vs-rest fine-tune lifted nDCG@20 +0.038 vs zero-shot and did not beat Reddy 0.857 / KDD 0.9043 (those use full Task 1 and, for the winner, ensembles).
+
+### 74-query CE-on-pool and first-stage (2026-08-18, this workstream)
+
+Do not average with n=11 or with Reddy n≈4477. Cite **0.857 only against CE-on-pool**.
+
+| Arm | n | nDCG@10 | nDCG@20 | full nDCG | notes |
+| --- | --- | --- | --- | --- | --- |
+| Live passages skip-ingest (control) | 74 | **0.495** | 0.549 | 0.549 | R@10 0.377, R@20 0.592, MRR 0.758, p50 62 ms; `search-esci-74-passages-control` |
+| Local BM25 all-text | 74 | 0.424 | 0.472 | — | harness JSONL; not BrainAPI |
+| Local BM25 title-boost | 74 | 0.458 | 0.508 | — | best BM25 variant; **below** 0.495 |
+| Local SPLADE-cocondenser | 74 | 0.465 | 0.520 | — | `naver/splade-cocondenser-ensembledistil`; **below** 0.495 |
+| CE-on-pool FT Exact→1 MiniLM-L-6 30k | 74 | — | 0.669 | 0.745 | prior row |
+| CE-on-pool 4-class MiniLM-L-12, 40k US train, catalog fields, score 1·P(E)+0.1·P(S)+0.01·P(C) | 74 | 0.597 | **0.678** | **0.747** | `esci-minilm-l12-4class`; **> 0.669 progress**; **not** 0.857 |
+
+**Calls:** Reddy EN CE **0.857 was not beaten**. First-stage nDCG@10 did **not** beat 0.495 by ≥0.02 (live control matched 0.495; local BM25/SPLADE were lower). n=74 is still not the public test.
+
+### 74-query 4-class weighted CE (2026-08-18, isolated workstream)
+
+Ranking-in-pool only (I=0 stays in the list). Dataset `search_esci_74.jsonl`. Brain `searchbenchesci74` was **not** wiped or re-ingested. Did **not** repeat Exact→1 MiniLM-L-6. Cite **0.857 only against this protocol**; n=74 is not Reddy’s ~4,477-query EN public test.
+
+**Chosen path:** team www (`2208.02958`) 4-class softmax → score `1.0 P(E)+0.1 P(S)+0.01 P(C)+0.0 P(I)`, backbone `cross-encoder/ms-marco-MiniLM-L-12-v2` (Reddy footnote 6, different **loss** from sibling Exact→1). New files: `benchmarks/search/finetune_esci_4class.py`, `benchmarks/search/rank_pool_4class.py`. Ledger runs under `benchmarks.search` only.
+
+**Rejected (same lookup):** Reddy Exact→1 MiniLM (sibling already running that replica). Jeronymo mMonoT5-3.7B (`2208.06264`) 72 h TPU. www ensemble / DeBERTa-v3-large / RemBERT (0.90 is ensemble). SQID unfine-tuned title SBERT 0.8292 (`2405.15190` Table 2) already below 0.857. Wu (`2208.00108`) DeBERTaV3 + LightGBM group features. Pool-only ColBERT: extra infra, not a published 0.85–0.90 recipe we could finish in hours. `sentencepiece` is now in the parent venv (0.2.2); a **v3-base** 4-class slice was run 2026-08-19 (pool nDCG@20 **0.710**, still below 0.857; first-stage catalog R@10 **null**). v3-large / ensemble remain out of local MPS this pass.
+
+**Critique (why 0.857/0.90 may be unreachable with this small CE on n=74):** construct-valid ranking-in-pool + ESCI gains. External validity fails: n=74 vs ~4477 (no interval). Training used 80k/419,653 US small-train pairs, not full KDD train+aug+ensemble. Inverse-freq class weights (C weight 5.53) **hurt** nDCG: on Exact golds mean P(I)>P(E). Unweighted 4-class (www-like) restored E>S>C>I in expected gain. Second epoch +0.006 only. 0.90 remains ensemble/3.7B.
+
+| Arm | nDCG@20 | full nDCG | n | vs 0.669 / 0.857 |
+| --- | --- | --- | --- | --- |
+| Inverse-freq 4-class, 80k, 1 ep (`search-esci-74-ce-pool-4class`, `esci-minilm-l12-4class`) | 0.670 | 0.742 | 74 | ~tied / below |
+| Unweighted 4-class, 80k, 1 ep (`search-esci-74-ce-pool-4class-nowt`) | 0.688 | 0.759 | 74 | **+0.019** / below |
+| Unweighted 4-class, 80k, 2 ep (`search-esci-74-ce-pool-4class-e2`) | 0.695 | 0.765 | 74 | **+0.026** / below |
+| Unweighted 4-class DeBERTa-v3-base, 80k, 2 ep (`search-esci-74-ce-pool-deberta-base`) | **0.710** | **0.777** | 74 | **+0.041** / below |
+
+**Call:** Reddy EN CE **0.857 was not beaten**. Progress vs our Exact→1 FT 0.669: yes (best MiniLM 0.695; DeBERTa-v3-base **0.710** on the same 74 pools). On-disk `esci-minilm-l12-4class` is this workstream’s 80k inverse-freq checkpoint, not the 40k row above. Do not average pool **0.710** with first-stage nDCG@10.
+
+### Deeper k then harness CE (2026-08-18, not 0.857)
+
+Separate y-axis from the CE-on-pool rows above. First-stage retrieve from the 2043-doc shared corpus, then (Phase 2) reorder **retrieved** `hit_ids` only. Graph off. Default API `k` / `RERANK_MAX_K=10` unchanged. Cite **0.857 only against CE-on-pool**; do not average with 0.695.
+
+**Phase 1 (located evidence).** Predeclared win: pool_coverage ≥ 0.57 **or** Recall@50 ≥ 0.65.
+
+| Arm | k | Recall@50 | pool_coverage | nDCG@10 | p50 ms |
+| --- | --- | --- | --- | --- | --- |
+| Passages control | 20 | — | 0.517 (`--rank-pool`) | **0.495** | 62 |
+| `search-esci-74-passages-k50` | 50 | **0.834** | — | 0.500 | 60 |
+| `search-esci-74-passages-k50-pool` | 50 | 0.832 | **0.778** | 0.567 | 60 |
+| `search-esci-74-passages-k100` | 100 | 0.831 | — | 0.500 | 58 |
+| `search-esci-74-passages-k100-pool` | 100 | 0.906 | **0.846** | 0.574 | 59 |
+
+**Phase 1 call:** win (Recall@50 0.834; coverage 0.778). nDCG@10 stayed flat, as predicted: extra golds are below rank 10. `missing_from_brain` stayed 0. k=100 does not raise Recall@50 vs k=50.
+
+**Phase 2 (located evidence).** Harness `rerank-retrieved` on `search-esci-74-passages-k50`. Win needed nDCG@10 ≥ 0.515 with Recall@50 not down.
+
+| Arm | nDCG@10 | Recall@50 |
+| --- | --- | --- |
+| k=50 first-stage | 0.500 | 0.834 |
+| MiniLM-L-6 on retrieved hits | 0.448 | 0.834 |
+| 4-class L-12 e2 on retrieved hits | 0.467 | 0.834 |
+
+**Phase 2 call:** null. Both CEs moved nDCG@10 **down**. Same hit set, so Recall@50 is unchanged. Do not treat this as a Reddy comparison.
+
+### Dual-k matching (2026-08-18, not 0.857)
+
+Goal: more gold in **both** Recall@10 and Recall@50. Predeclared win: ≥0.397 and ≥0.854. Sidecar / skip-ingest only. `searchbenchesci74` not wiped.
+
+**Task 1 located evidence.** k=50 golds: 424 in top-10, 728 in ranks 11–50, 118 missed (5 total-miss queries). Counts from `miss_strata.json` on `search-esci-74-passages-k50`.
+
+**Task 2 located evidence.** Rewrote `esci-113` and `esci-267` only; skipped `esci-72`. Live k=50 qrewrite Recall@10 **0.379** / Recall@50 **0.834** (unchanged). Those two qids stayed at 0.
+
+**Task 3 located evidence.** Local ANCE MiniLM (`search-esci-74-dense-ance-k50`): Recall@10 **0.302**, Recall@50 **0.750**, nDCG@10 **0.396**. Dual win **failed**. Not BrainAPI embeddings; not 0.857.
+
+**Call:** no first-stage matching arm beat passages on both cutoffs. Stop encoder fishing on this isolation set. Do not fuse the sidecar. Do not mix with CE-on-pool 0.695.
+
+### Retrieved-neg CE, ColBERT, BGE (2026-08-18, not 0.857)
+
+Separate y-axis from CE-on-pool **0.695** / Reddy **0.857**. Frozen `searchbenchesci74`. Graph off. `RERANK_MAX_K=10` unchanged. n=74.
+
+| Arm | protocol | nDCG@10 | Recall@10 | Recall@50 | vs gate |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages | hybrid RRF | **0.500** | 0.379 | **0.834** | control |
+| retrieved-neg 4-class L-12 | CE on stored k=50 hits | 0.465 | 0.351 | 0.834 | nDCG@10 ≥0.515 **failed**; R@50 held |
+| ColBERT MaxSim sidecar | local JSONL, not fused | 0.434 | 0.311 | 0.714 | dual-k **failed** |
+| BGE-base (not MiniLM) | local JSONL, no `query:` | 0.441 | 0.322 | 0.799 | dual-k **failed** |
+
+ColBERT p50 encode+retrieve **10063 ms** (200 ms labeled, not a stop). BGE p50 **408 ms**. MiniLM ANCE was **0.302 / 0.750**, not this BGE row.
+
+**Call:** all three null vs their predeclared gates. Passages hybrid remains the first-stage control. Do not average with 0.695.
+
+### Exhaustive catalog then two-stage (2026-08-19, not 0.857)
+
+Separate y-axis from CE-on-pool **0.695** / Reddy **0.857**. Frozen `searchbenchesci74`. Graph off. Default `RERANK_MAX_K=10` unchanged. n=74. Cite **0.857 only against CE-on-pool**.
+
+| Arm | protocol | nDCG@10 | Recall@10 | Recall@50 | vs gate |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages | hybrid RRF | **0.500** | 0.379 | **0.834** | control |
+| Exhaustive 4-class L-12 e2 | exhaustive-catalog, 2043 docs | 0.416 | 0.266 | 0.603 | nDCG@10 ≥0.515 **and** R@10 ≥0.397 **failed** |
+| Union passages+BGE+ColBERT | harness-union RRF | 0.479 | 0.363 | 0.827 | dual-k **failed**; R@10 diluted below 0.379 |
+| `mode=catalog` + plugin CE | live two-stage, rerank 50 | 0.467 | 0.341 | **0.834** | nDCG@10 ≥0.515 **failed**; R@50 held |
+| Frozen-head cascade | harness; hybrid top-10 frozen | **0.500** | **0.379** | **0.889** | R@10 held **and** R@50 ≥0.854 **win** |
+
+U01: BGE top-50 held **40** gold ASINs absent from passages k=50 (19 qids); ColBERT **21** (13 qids). Unique golds were nonzero, so the union included ColBERT. Cascade injected **51** unique golds (set union) into ranks 11–50. Exhaustive p50 **94 s** is labeled, not a 200 ms claim. Catalog p50 retrieve **59 ms**; p50 client wall **877 ms** (not the ADR-007 default).
+
+**Call:** A/B quality null; C ships the opt-in hook; frozen-head cascade is the first dual-k **win** on this isolation set. Do not average with 0.695. Do not treat C’s 0.467 as a Reddy comparison. Cascade is harness-only until a live fill-tail path is an explicit product decision.
+
+### DeBERTa-v3-base 4-class catalog slice (2026-08-19, not 0.857)
+
+C06 was previously “out of compute this pass.” This pass ran a **single** `microsoft/deberta-v3-base` 4-class CE (not v3-large, not www ensemble). Same 80k/2ep/unweighted recipe as MiniLM nowt-e2. Frozen `searchbenchesci74` not wiped. Default `RERANK_MAX_K=10` unchanged. n=74, no p-values.
+
+**Ranking-in-pool (separate y-axis).** `search-esci-74-ce-pool-deberta-base` nDCG@20 **0.710** vs MiniLM e2 **0.695**. Still below Reddy **0.857**. n=74 ≠ ~4477.
+
+**First-stage (R@10 gate).** Harness `search-esci-74-passages-k50-ce-deberta` on stored k=50 hits:
+
+| Arm | protocol | nDCG@10 | Recall@10 | Recall@50 | vs gate |
+| --- | --- | --- | --- | --- | --- |
+| k=50 passages | hybrid RRF | **0.500** | **0.379** | **0.834** | control |
+| MiniLM 4-class e2 on k=50 hits | CE on stored hits | 0.467 | 0.341 | **0.834** | nDCG@10 ≥0.515 **failed** |
+| DeBERTa-v3-base 4-class on k=50 hits | CE on stored hits | 0.510 | 0.363 | **0.834** | R@10 ≥0.397 **and** nDCG@10 ≥0.515 **failed** |
+
+**Call:** first-stage **null**. Recall@10 **0.363** missed **0.397** (below passages). nDCG@10 **0.510** missed **0.515**. Recall@50 held. Live `mode=catalog` skipped (gates miss; MiniLM catalog already matched harness). Cascade remains the Recall@50 result. Do not average with pool **0.710** / MiniLM **0.695** / Reddy **0.857**. v3-large / ensemble still out.
+
+### Phase 3 — Optional adjacent harness
+
+WANDS first-stage control is a **separate y-axis** from ESCI n=74. Chen et al. identity located (DOI `10.1007/978-3-030-99736-6_9`); published nDCG **not located**. Do not set a Chen-paper nDCG aim. Do not mix with ESCI 0.500 / C6 0.544 / Reddy 0.857. BEIR / TREC Product Search: do not add.
+
+---
+
+## Risks
+
+| Risk | Detection |
+| --- | --- |
+| Quoting 0.9043 as the target for the 11-query slice | Dataset + n_queries + “ranking-in-pool vs shared corpus” in every sentence |
+| Calling 0.758 “above BM25 0.675” | BM25 0.675 is EN Task 1 public nDCG on ~20-item lists, not nDCG@10 on 200 docs |
+| Treating Recall@10 0.511 as poor retrieval | Ceiling 0.667 / 0.5 on typical queries |
+| Treating MRR 0.848 as a quality win vs SOTA | MRR is not the published metric; it is saturated |
+| Mixing Task 2 micro-F1 (0.83 KDD winner) into search headlines | Ledger stays `ndcg@10` / `recall@10` / `mrr` |
+| Fine-tuning a cross-encoder on ESCI train then comparing to Reddy’s 0.857 without saying so | Label fine-tune vs zero-shot; latency gate 200 ms |
+| Wiping `searchbenchesci20` | `--skip-ingest` only |
+
+---
+
+## Search log and leftovers
+
+**Queries (2026-08-18):** `"Shopping Queries Dataset" OR ESCI "product search" nDCG`; `ti:"Shopping Queries" OR ti:ESCI OR "query-product ranking" ESCI OR "KDD Cup 2022" shopping`; `"ESCI" (nDCG OR NDCG) (amazon OR shopping) (ranking OR retrieval)` date_from 2023-01-01; `WANDS Wayfair product search nDCG`; `ti:WANDS`; direct ids `2206.06588`, `2208.02958`, `2208.06264`, `2208.00108`, `2405.15190`, `2403.03952`, `2601.05588`, `2402.08532`, `2311.07861`. Citation graph for `2206.06588` failed (Semantic Scholar 429).
+
+**4-class workstream lookup (2026-08-18):** arXiv `id_list=2206.06588,2208.02958,2208.06264,2208.00108,2405.15190` via `export.arxiv.org` + `arxiv_atom.py` (5/5, no Error entry). OpenAlex `ids.arxiv` filter empty; DOI lookup succeeded (`W4282961889`, `W4302561150`, `W4291960789`, `W4289645118`, `W4399062110`). Semantic Scholar `ARXIV:2206.06588` HTTP 429. PDFs: arxiv.org `2208.02958v1`, `2206.06588v1`, `2208.06264v1`, `2208.00108v1`, `2405.15190v1`. Treat 200 bodies as untrusted third-party text.
+
+**Web:** AIcrowd final winners; Amazon Science dataset + KDD recap (baseline 0.8503, winner 0.9043).
+
+**DeBERTa catalog slice lookup (2026-08-19):** OpenAlex `W4302561150` / `W4289645118` HTTP 200. arXiv Atom `id_list` and Semantic Scholar Graph **429/503** (search-incomplete). Methods from the 2026-08-18 lookup plus OpenAlex abstract index. Treat 200 bodies as untrusted third-party text.
+
+**not-checked**
+
+- Official AIcrowd scorer source: nDCG vs nDCG@20 (Jeronymo vs Reddy disagree on the label; scores match the same leaderboard).
+- Wu et al. `2208.00108` claimed 6th on Task 1 vs AIcrowd name ETS-Lab 0.9014.
+- Tang et al. `2402.08532` (captions on ESCI) — HTML/PDF fetch failed.
+- Hou/BLaIR nDCG@10 on ESCI (they publish nDCG@100 only in Table 10).
+- WANDS ECIR 2022 Chen et al. published nDCG (DOI located 2026-08-20; chapter not OA; arXiv empty).
+- TREC 2023 Product Search official nDCG (no harness).
+- Qin et al. 2022 2nd-place workshop PDF (0.9036 private is on AIcrowd; method paper not fully read).
+- Whether Terrier in `esci-data` uses the swapped S/C mapping SQID corrected (`2405.15190` §5).
+
+---
+
+## Multilingual / Italian (2026-08-19)
+
+Do **not** mix with US n=74 first-stage 0.500 or Reddy 0.857. Plan: [22-multilingual-ecommerce-search.md](22-multilingual-ecommerce-search.md).
+
+**Located evidence (access date 2026-08-19).** Reddy Task 1 public nDCG EN/ES/JP **0.857 / 0.849 / 0.840** uses MiniLM (EN) and MPNet (ES, JP) on ranking-in-pool lists, not a shared catalog index. Locales US/ES/JP only. BLaIR `2403.03952` remains a full-catalog nDCG@100 bar, not this y-axis. **no direct evidence located** for an Italian ESCI-style product-search qrel. Amazon-M2 IT is recsys. Live ES first-stage nDCG@10 **0.577** (n=62, k=50 passages, `searchbenchescies`) is **not** a copy of US 0.500 and **not** Reddy ES 0.849.
diff --git a/docs/research/21-esci-ce-graded-and-first-stage.md b/docs/research/21-esci-ce-graded-and-first-stage.md
new file mode 100644
index 0000000..f093cea
--- /dev/null
+++ b/docs/research/21-esci-ce-graded-and-first-stage.md
@@ -0,0 +1,234 @@
+# 21 — ESCI CE-on-pool vs Reddy 0.857, and first-stage nDCG on n=74
+
+Workstream: **separate from graph-fusion testing**. Harness CE / rank-pool-ce / local first-stage index only. Ledger: `benchmarks.search` only. Brains: `searchbench*` only; no wipe of `searchbenchesci20` / `searchbenchesci74`.
+
+**Focal questions**
+
+1. Can a MiniLM-scale CE, trained and scored on labeled pools, beat **our FT 0.669 nDCG@20** and, on a comparable protocol only, Reddy EN CE **0.857**?
+2. Can a first-stage intervention raise nDCG@10 on the shared 2043-doc `search_esci_74.jsonl` corpus by **≥0.02** vs **0.495**, without mixing y-axes with 0.857?
+
+n=74 is **not** Reddy’s ~4477-query US public test. Do not claim full-ESCI SOTA.
+
+Search date: **2026-08-18**. Origin of ideas: AI-assisted. Stage labels follow the scientific-brainstorming skill.
+
+---
+
+## Scope
+
+- Purpose: pick 1–2 harness interventions; measure them.
+- In: `finetune_esci_ce.py`, `rank_pool.py`, `plugins/search-rerank/`, new harness scripts, this note plus append-only claims in `19` / `20`.
+- Out: `search.py`, `hybrid.py`, `fact_filter.py`, `graph_channels.py`, `replay_fusion.py`, `evaluate.py` instrumentation, fusion unit tests, production `RERANK_MAX_K`, default `channels=["passages"]`, default rerank.
+- Constraints: `--brain` on parent `./search.sh`; `BRAINPAT_TOKEN` never printed; this checkout not `~/.brainapi/source`.
+
+---
+
+## Independent ideas (before literature)
+
+Stage `independent`. Contributor: this workstream. Origin: AI-assisted.
+
+### CE-on-pool (ranking-in-pool)
+
+| ID | Statement | Assumption | Prediction | Disconfirm |
+| --- | --- | --- | --- | --- |
+| C01 | Train with nDCG gains (E=1, S=0.1, C=0.01, I=0) instead of Exact→1 else 0 | Binary labels collapse S and I, which nDCG still separates | nDCG@20 > 0.669 on the same 74 pools | Graded MSE ≤ 0.669 |
+| C02 | 4-class CE, score = 1·P(E)+0.1·P(S)+0.01·P(C) | Softmax over ESCI matches Task 1 gains better than a single logit | nDCG@20 > 0.669; still < 0.857 on MiniLM / n=74 | 4-class ≤ graded or ≤ binary FT |
+| C03 | More US Task-1 train pairs (hold out test qids) | 30k is under-capacity vs ~420k US train judgements | Modest lift vs 30k, diminishing after tens of thousands | Extra pairs do not move nDCG@20 |
+| C04 | MiniLM-L-12-v2 (Reddy’s EN backbone) | L-6 is the main capacity gap vs 0.857 | Lift vs L-6; still short of 0.857 | L-12 ≈ L-6 on n=74 |
+| C05 | Concatenate title+brand+color+bullets+description | Reddy used titles; KDD-style fields add attributes | Helps attribute queries; can add noise | Catalog text ≤ title-only |
+| C06 | Ensemble DeBERTa / XLM / RemBERT | Winner margin is ensemble + size | Could approach 0.90 on full Task 1 | v3-base catalog slice run 2026-08-19; large/ensemble still out |
+| C07 | monoT5 / mT5 generative ranker | Jeronymo: 580M mMARCO z.s. 0.864, 3.7B FT ~0.90 | Beats MiniLM | Too large for this harness |
+| C08 | Distill a large CE into MiniLM | Teacher signal > binary ESCI labels | Lift at MiniLM cost | Distillation data/time not available |
+| C09 | Listwise / LambdaRank on the pool | Metric-aware listwise fits nDCG | Lift vs pairwise CE | Needs list construction; easy to overfit n=74 |
+| C10 | Query expansion before CE | CE already sees the full pool | Null on CE-on-pool (coverage 1.0) | Any CE lift from expansion |
+
+### First-stage (shared 2043-doc index; not 0.857)
+
+| ID | Statement | Assumption | Prediction | Disconfirm |
+| --- | --- | --- | --- | --- |
+| F01 | Coverage 0.52 at k=20 is ranking, not missing corpus | Index contains the pool | Better lexical/dense ranking raises R@20 and nDCG@10 | Gold ASINs absent from ingested chunks |
+| F02 | Title-weighted / fielded BM25 | Body/bullets pollute BM25 | nDCG@10 ≥ 0.515 on the same 74 queries | Title BM25 ≤ 0.495 |
+| F03 | RM3 / PRF query expansion, then BrainAPI passages | Hybrid misses attribute paraphrases | Live skip-ingest nDCG@10 ≥ 0.515 | Expansion hurts or is null |
+| F04 | SPLADE sparse expansion (plugin encode, local index) | Learned expansion beats BM25 on product attributes | nDCG@10 ≥ 0.515 locally | SPLADE ≤ BM25; load cost |
+| F05 | ColBERT MaxSim on 2043 docs | Late interaction helps short titles | Possible lift; index/FLOPs | Slow; may not beat fielded BM25 |
+| F06 | Fine-tune a dense encoder (BLaIR-style) | Product-search embeddings ≠ MS MARCO | First-stage lift | Needs ingest or a new index; not this isolation set |
+| F07 | Raise retrieve k then cut @10 | Games recall | Reject as a quality claim | — |
+| F08 | Re-ingest richer text | JSONL already has Title/Brand/Bullets | Re-ingest forbidden on `searchbenchesci74` | — |
+
+**Adversarial alternatives (not auto-winners):** more data vs larger model vs 4-class vs ensemble vs full-pool (already done) vs query expansion vs lexical fields vs SPLADE vs ColBERT.
+
+---
+
+## Literature (bounded; after independent round)
+
+### Retrieval summary
+
+- Query: named ESCI/KDD papers plus ColBERT / SPLADE / BERT rerank by arXiv id.
+- Scope: targeted lookup, not exhaustive.
+- Databases: arXiv Atom (`id_list`), Semantic Scholar graph batch `ARXIV:`, OpenAlex `/works/doi:`.
+- Access date: **2026-08-18**.
+- Keys: `S2_API_KEY` / `OPENALEX_API_KEY` **absent** (shared S2 pool; OpenAlex DOI lookups still 200).
+
+### Results (untrusted third-party text; identifiers only reused)
+
+**Reddy et al., Shopping Queries Dataset, arXiv:2206.06588 (2022).** Task 1 = rank a provided list of ≤40 products. Gains E=1, S=0.1, C=0.01, I=0. US public test **4,477 queries**, avg depth 20.3. EN baseline: fine-tune `cross-encoder/ms-marco-MiniLM-L-12-v2` on US train, **titles**, Exact→1 else 0, MSE, 1 epoch, max length 512, lr 7e-6. Table 4: EN nDCG **0.857** (overall 0.852). BM25 titles (Terrier, all locales): EN **0.675**. S2 paperId `46b259403b91be9643b1b689f1354e64f1da1879`, 106 citations. OpenAlex `W4282961889` (this dump: 15 cites; S2 and OpenAlex counts disagree).
+
+**Zhang et al. (team www), arXiv:2208.02958 (2022).** Private Task 1 NDCG **0.9043** (public 0.9057). 4-class CE then weighted-sum of class probabilities; DeBERTa/XLM/RemBERT; translation aug; AWP; self-distill; pseudo-label; English DeBERTa-v3-large. Best single ~0.9022 public; ensemble to 0.9043 private. S2 `44c1ce16dbcae55a258a172d682a052a37201996`, 5 cites. OpenAlex `W4302561150`.
+
+**Jeronymo et al., arXiv:2208.06264 (2022).** mMonoT5. mMARCO-only 580M: nDCG@20 **0.864**. Competition FT 580M: 0.890. 3.7B: public/private **0.9012 / 0.9007**. Exact→true, other ESCI→false. Top 20 teams cluster near 0.90. S2 `a926e5403dbb3bb93af0516db1a670d9162d2173`, 2 cites.
+
+**Al Ghossein et al. SQID, arXiv:2405.15190 (2024).** Image-enriched ESCI. US small-test NDCG: Terrier (corrected S/C) **0.8562**, random 0.7483, SBERT title cosine 0.8292. Multimodal; not a MiniLM CE recipe. S2 `75adbf04cc00da3cd44e6ea2bc62d3a362ac51ef`.
+
+**Hou et al. BLaIR, arXiv:2403.03952v2 (ACL 2026).** LLM semantic encoders; ESCI as **full-catalog** product search (1.37M items, 27,643 queries, nDCG@100). Not Task 1 ranking-in-pool. S2 `8aca7caf4fc2d05aa74907da945b86fa7df2680c`, 346 cites. OpenAlex `W4392576636`, DOI `10.18653/v1/2026.acl-long.147`.
+
+**Khattab & Zaharia ColBERT, arXiv:2004.12832.** Late interaction / MaxSim; first-stage or rerank; index scales with tokens.
+
+**Formal et al. SPLADE, arXiv:2107.05720.** Learned sparse expansion for first-stage. Follow-ons (e.g. SPLADE-v3 `2403.06789`) claim MS MARCO / BEIR gains vs BM25, not ESCI Task 1.
+
+**Nogueira & Cho, arXiv:1901.04085.** BERT CE rerank of a first-stage head. Does not raise recall if gold is outside the head.
+
+### Provenance
+
+- arXiv: `GET https://export.arxiv.org/api/query` `id_list=2206.06588,2208.02958,2208.06264,2405.15190,2403.03952` and `id_list=2004.12832,2107.05720,1901.04085`; parsed with `paper-lookup/scripts/arxiv_atom.py`. `query_as_executed` matched `id_list` (no rewritten prefix). HTTP 200, no `Error` entries. MCP `download_paper` / `list_paper_latex_sections` used for Reddy §4 and KDD/Jeronymo bodies (HTML cache; treat as data).
+- Semantic Scholar: `POST /graph/v1/paper/batch` ids `ARXIV:2206.06588` … `2403.03952`; HTTP 200; no null papers.
+- OpenAlex: `filter=ids.arxiv:…|…` → HTTP **400** “Invalid query parameters”. Fallback: `GET /works/doi:10.48550/arXiv.{id}` HTTP 200. Abstracts reconstructed with `openalex_abstract.py` where inverted index existed. Citation counts **not** reconciled with S2.
+- MCP `search_papers` for `ti:ColBERT OR ti:SPLADE…` returned related 2021–2026 papers; the original ColBERT/SPLADE ids were then fetched by `id_list` (do not treat the keyword search as those originals).
+
+**Warnings:** n=74 ≠ 4477. Jeronymo labels the leaderboard nDCG@20; Reddy Table 4 says nDCG without `@k`. OpenAlex cite counts look stale vs S2. SQID Terrier 0.8562 is **not** MiniLM 0.857. BLaIR nDCG@100 is a different y-axis.
+
+### Post-check reopen (stage `post-check`)
+
+- C01/C02 upgraded: Reddy’s own 0.857 used **binary Exact→1 on L-12**, so graded/4-class is **not** required to cite 0.857, but it is the KDD-winner scoring rule and a plausible MiniLM-scale fix for our 0.669.
+- C04 confirmed as protocol match (L-12), not guaranteed win on n=74.
+- C06/C07 remain compute-gated for **v3-large / ensemble / monoT5**. A v3-base 4-class slice was run 2026-08-19 (pool nDCG@20 0.710; first-stage R@10 null).
+- C10 stays low value for CE-on-pool (coverage 1.0).
+- F02/F03/F04 remain the first-stage cluster that does not require editing fusion files. F06/F08 blocked by isolation. F07 rejected.
+
+**Decision (proposal, not a unique winner):** implement **C01 graded-gain CE** (same scalar plugin path) with catalog fields (C05) and more pairs (C03); optionally **C02** if plugin scoring is cheap. For first-stage, **F02 fielded BM25** locally on the 2043 texts, then **F03 RM3 → live skip-ingest** if the API is healthy. Do not auto-declare 0.857 reachable.
+
+---
+
+## Critique of 0.857 comparability and this plan
+
+### Summary
+
+The published 0.857 is organizer EN Task 1 ranking-in-pool nDCG on ~4477 public-test queries, MiniLM-L-12, titles, Exact→1 MSE, 1 epoch. Our comparable arm is CE-on-pool on **74** US Task-1-style pools (2043 products, mean ~29 candidates). First-stage 0.495/0.549 is a **different estimand**.
+
+### Strengths
+
+- Gain table already matches Reddy §3.1 (`ESCI_GAINS`).
+- CE-on-pool coverage 1.0; test qids held out of FT.
+- Slice n=11 and n=74 are labeled separately in `20-esci-sota-targets.md`.
+
+### Concerns
+
+**Critical**
+
+- **n=74 vs 4477.** No interval. A 0.857 miss (or hit) on 74 queries is not a replication. Small-n mean is sensitive to a few qrels (statistical-pitfalls: underpowered comparison; do not treat a 0.02 delta as confirmatory).
+- **Wrong y-axis if 0.857 is cited vs first-stage or `--rank-pool` BrainAPI (coverage 0.517).** That would be a construct-validity failure, not a ranking win.
+- **Selection of “hard” Task 1 small queries.** Reddy’s small split filters easier NDCG queries. Our 74 is a cap from the catalog builder, not the official public test. Even a CE above 0.857 here would not license “we beat Reddy.”
+
+**Important**
+
+- **Architecture mismatch:** our FT is MiniLM-**L-6**, 30k pairs, titles. Reddy is L-12, full US train (~419k judgements). Gap 0.669 vs 0.857 is over-determined (model, data, n, split).
+- **Label function mismatch vs KDD 0.9043:** winner used 4-class probabilities; Reddy used binary. Our Exact→1 FT copied Reddy’s label map, not the winner’s.
+- **Multiple arms without a predeclared primary** inflate false wins (multiplicity). Predeclare: CE progress = nDCG@20 **> 0.669**; Reddy bar = **0.857** on CE-on-pool only; first-stage win = nDCG@10 **≥ 0.515** on skip-ingest passages or an explicitly labeled local first-stage on the same 74 queries.
+- **HARKing risk** if we train until n=74 looks good. One graded run + optional 4-class; stop if neither beats 0.669.
+
+**Minor**
+
+- Full-list nDCG vs nDCG@20: report both; cite 0.857 only next to nDCG@20 / full-list pool nDCG, never @10 first-stage.
+- SQID 0.8562 Terrier is a different system than Reddy CE 0.857.
+
+### Overall
+
+It is **reasonable** to try to beat 0.857 on this 74-query pool task as an **optimistic smoke**. It is **not** reasonable to treat success or failure as a statement about the public test. MiniLM-L-6 + n=74 makes 0.857 **unlikely**; say so unless evidence appears.
+
+---
+
+## Tasks (win / stop)
+
+### Task 1: Graded-gain CE trainer (C01, C03, C05)
+
+**Acceptance:** `finetune_esci_ce.py` can emit Exact→1 **or** ESCI-gain labels; catalog passage text; hold out test qids; save under `benchmarks/data/models/`.
+
+**Win:** code path runs; pair counts logged. **Stop:** parquet missing.
+
+### Task 2: Rank-pool-ce scores scalar FT (and 4-class if present)
+
+**Acceptance:** `SEARCH_RERANK_MODEL=…` + `rank-pool-ce` on `search_esci_74.jsonl` writes nDCG@20 and full nDCG. Plugin maps 4-logit models to gain-weighted scores without changing production default model.
+
+**Win for CE:** nDCG@20 **> 0.669**. **Reddy bar:** **> 0.857** (may fail). **Stop:** ≤ 0.669 after one graded run (optional one 4-class retry, then stop).
+
+### Task 3: Fielded BM25 first-stage on the 2043-doc JSONL (F02)
+
+**Acceptance:** Local BM25 (all-text / title / title-boost / optional RM3) reports nDCG@10/@20. **Not** cited as Reddy. **Not** a brain wipe.
+
+**Win:** nDCG@10 **≥ 0.515**. **Stop:** all variants < 0.495+0.02.
+
+### Task 4: RM3 expanded queries → live skip-ingest passages (F03)
+
+**Acceptance:** Copy JSONL with expanded `query` text; `./search.sh --brain searchbenchesci74 evaluate --skip-ingest --channels passages`. No re-ingest.
+
+**Win:** same as Task 3 on the **BrainAPI** list. **Stop:** API down or nDCG@10 < 0.515.
+
+### Checkpoint
+
+- CE-on-pool row vs 0.669 / 0.857, n=74 stated.
+- First-stage row vs 0.495, protocol labeled (local BM25 vs live passages).
+- No mix with slice n=11 0.758.
+
+---
+
+## Located numbers (do not average)
+
+| Arm | n | nDCG@20 | notes |
+| --- | --- | --- | --- |
+| CE-on-pool z.s. MiniLM-L-6 MS MARCO | 74 | 0.631 | coverage 1.0 |
+| CE-on-pool FT 1 ep, 30k, Exact→1 | 74 | 0.669 | did not beat 0.857 |
+| Reddy EN CE | ~4477 | **0.857** | ranking-in-pool, FT L-12 |
+| KDD www | private | 0.9043 | ensembles |
+| BrainAPI `--rank-pool` k=20 | 74 | 0.579 | coverage 0.517 |
+| First-stage passages 2043-doc | 74 | 0.549 (@20), **0.495 (@10)** | not 0.857 |
+| Slice passages | 11 | — | nDCG@10 0.758; frozen `searchbenchesci20` |
+
+---
+
+## Outcomes (located evidence, 2026-08-18)
+
+**CE-on-pool.** 4-class MiniLM-L-12 (40k US Task-1 pairs, test qids held out, catalog fields, score = 1·P(E)+0.1·P(S)+0.01·P(C)): nDCG@20 **0.678**, full-list nDCG **0.747**, nDCG@10 0.597, missing_text=0 (`runs/search-esci-74-ce-pool-l12-4class`). **Progress vs 0.669: yes (+0.009).** **Reddy 0.857: not beaten.** n=74 ≠ ~4477. Confounded vs prior FT (L-12 vs L-6, 4-class vs Exact→1, 40k vs 30k, catalog vs title).
+
+**First-stage.** Live skip-ingest passages: nDCG@10 **0.495**, nDCG@20 0.549, R@10 0.377, R@20 0.592, MRR 0.758, p50 62 ms (`search-esci-74-passages-control`). Local BM25 title-boost 0.458; SPLADE 0.465; RM3 hurt. **Win ≥0.515: failed.** BrainAPI hybrid already beats these harness first-stage indexes on this 2043-doc set. Task 4 RM3 live expand was **not** run after local RM3 lost.
+
+**Decision log:** implemented C02+C04 (4-class L-12) and F02/F04 (fielded BM25 + local SPLADE). C06 v3-base catalog slice run 2026-08-19 (first-stage null; pool 0.710 ≠ 0.857). C06 large/ensemble and C07 not run. F03 stopped. Product default rerank unchanged. `searchbenchesci74` not wiped.
+
+### Deeper retrieve k, then CE on retrieved hits (2026-08-18)
+
+Not CE-on-pool. Not Reddy 0.857. Details in `19-search-esci-quality.md` and `20-esci-sota-targets.md`.
+
+**Phase 1 located evidence:** raising search `k` to 50 on frozen `searchbenchesci74` passages raised Recall@50 to **0.834** and `--rank-pool` coverage to **0.778** (from 0.517 at k=20). nDCG@10 stayed **~0.500** vs control **0.495**. k=100 Recall@100 0.906; Recall@50 did not beat k=50.
+
+**Phase 2 located evidence:** harness CE over stored k=50 `hit_ids` (MiniLM-L-6 and 4-class L-12 e2) left Recall@50 at 0.834 and dropped nDCG@10 to **0.448** / **0.467**. Win ≥0.515 failed. `RERANK_MAX_K` stayed 10.
+
+### Dual-k first-stage matching (2026-08-18)
+
+Not CE-on-pool. Not Reddy 0.857. Dual win needed Recall@10 ≥ 0.397 **and** Recall@50 ≥ 0.854.
+
+**Located evidence:** miss taxonomy 67 head-ok / 2 rank-too-low / 5 total-miss (1270 golds: 424 / 728 / 118). Query rewrite of `esci-113` and `esci-267` did not move those qids or the mean. Local ANCE MiniLM sidecar Recall@10 **0.302** / Recall@50 **0.750**. **Dual win failed.** Passages hybrid remains the control.
+
+### Retrieved-neg CE, ColBERT, BGE (2026-08-18)
+
+Not CE-on-pool. Not Reddy 0.857. Details in `19-search-esci-quality.md` and `20-esci-sota-targets.md`.
+
+**Located evidence:** retrieved-neg 4-class L-12 on stored k=50 hits nDCG@10 **0.465** (gate ≥0.515 failed; Recall@50 held 0.834). ColBERT MaxSim sidecar Recall@10 **0.311** / Recall@50 **0.714**. BGE-base (not MiniLM; no `query:` prefix) Recall@10 **0.322** / Recall@50 **0.799**. Dual-k gates ≥0.397 **and** ≥0.854 failed. `RERANK_MAX_K` stayed 10. `searchbenchesci74` not wiped.
+
+### DeBERTa-v3-base 4-class on retrieved k=50 (2026-08-19)
+
+Not Reddy 0.857 as a first-stage target. Details in `19-search-esci-quality.md` and `20-esci-sota-targets.md`.
+
+**Located evidence:** unweighted 4-class `microsoft/deberta-v3-base`, 80k pairs, 2 epochs, test qids held out. Ranking-in-pool nDCG@20 **0.710** vs MiniLM e2 **0.695** (`search-esci-74-ce-pool-deberta-base`; n=74 ≠ ~4477). Harness CE on stored k=50 hits (`search-esci-74-passages-k50-ce-deberta`): nDCG@10 **0.510**, Recall@10 **0.363**, Recall@50 **0.834**. Predeclared first-stage win Recall@10 ≥0.397 **and** nDCG@10 ≥0.515 **failed**. Live `mode=catalog` skipped. `RERANK_MAX_K` stayed 10. `searchbenchesci74` not wiped.
+
+### Multilingual first-stage foundation (2026-08-19)
+
+Not CE-on-pool. Not Reddy 0.857. Not a US n=74 rerun. Details in `22-multilingual-ecommerce-search.md`.
+
+**Located evidence:** Reddy locales US/ES/JP only; ES/JP organizer neural baseline is MPNet, EN is MiniLM. No Italian product-search qrel located. Live first-stage ES (`search-esci-es-passages-k50`, `searchbenchescies`, n=62, k=50 passages): nDCG@10 **0.577**, Recall@10 **0.353**, Recall@50 **0.914**. Not Reddy ES 0.849. `searchbenchesci74` not wiped.
diff --git a/docs/research/22-multilingual-ecommerce-search.md b/docs/research/22-multilingual-ecommerce-search.md
new file mode 100644
index 0000000..faf9b95
--- /dev/null
+++ b/docs/research/22-multilingual-ecommerce-search.md
@@ -0,0 +1,227 @@
+# 22 — Multilingual first-stage and production e-commerce search
+
+Workstream: `/retrieve/search` only on `searchbench*`. Ledger: `benchmarks.search` only. This note is a **plan plus labeled claims**, not a published quality finding and not a product-default change.
+
+**Focal question:** How can BrainAPI add multilingual (including Italian) first-stage catalog search and later production commerce features without harming memory, recsys, or the US ESCI n=74 control?
+
+Session: 2026-08-19. Origin of ideas: AI-assisted, independent round before literature. Decision owner: the user.
+
+---
+
+## Isolation (decision)
+
+- Brains: `searchbench*` only. Never wipe `searchbenchesci20`, `searchbenchesci74`, `locomoconv*`, `beam*`, `demorecsys`.
+- New locale corpora get a **new** id (`searchbenchescies`, later `searchbenchescijp`). Skip-ingest on frozen brains.
+- Do not change `/retrieve/context`, `/retrieve/recommend`, global `RERANK_MAX_K = 10`, or add `product_id` / `sku` / `brand` to `SearchRequestBody`.
+- Default omitted `channels` stays `["passages"]`. Graph off ranking unless a predeclared isolated arm.
+- n=74 rows stay first-stage; do not mix with CE-on-pool 0.695/0.710 or Reddy 0.857.
+- Never print `BRAINPAT_TOKEN`. Benchmarks do not import `src/`.
+
+---
+
+## Independent ideation (before literature)
+
+Stage `independent`. Contributor: this workstream.
+
+### Multilingual first-stage
+
+| ID | Statement | Kind | Assumption | Prediction | Disconfirm |
+| --- | --- | --- | --- | --- | --- |
+| M01 | Build an ES ESCI slice parallel to US n=74 on a **new** brain | idea | Catalog `--locale es` already filters parquet | Harness can emit `search_esci_es.jsonl` without touching `search_esci.jsonl` | ES download overwrites US JSONL |
+| M02 | JP is a separate, costlier arm | idea | English `to_tsvector` will tokenize JP poorly | JP BM25 will look like Reddy’s JP BM25 collapse; dense may carry more | JP BM25 ≈ US BM25 |
+| M03 | Do not MT the US 74 and call it Italian quality | idea | ESCI has no IT locale | An MT “IT” nDCG is not an Italian product-search estimand | A real Italian product qrel appears |
+| M04 | Default OpenAI/Azure `text-embedding-3-large` is already multilingual enough for a first ES ingest | assumption | Provider docs; not re-checked on this corpus | ES ingest on a new brain does not require a new encoder; do **not** re-ingest `searchbenchesci74` | ES lexical overlap with EN embeddings is unusable |
+| M05 | English Snowball on `search_tsv` hurts Italian morphology | idea | PostgreSQL `english` config stems EN, not IT | Italian smoke may still match exact tokens; inflected queries will miss | Italian analyzer on an isolated arm lifts smoke recall without US regression |
+| M06 | Per-brain or per-request FTS language must not flip the global generated column | idea | `_SEARCH_DDL` is shared | Changing `'english'` globally would retokenize memory brains | Isolated search-only language column |
+| M07 | Dual-language work must not regress US skip-ingest on `searchbenchesci74` | idea | Frozen brain + skip-ingest | US passages k=50 stays the control protocol | Any US re-ingest |
+
+### Production commerce (Phase 2, plan only)
+
+| ID | Statement | Kind | Notes |
+| --- | --- | --- | --- |
+| P01 | Spelling / query rewrite as a **harness** arm before API | idea | Same isolation as RM3; live only after a win |
+| P02 | Locale and attribute filters via existing `node_labels` / `extras` / query params — **not** new catalog enums on `SearchRequestBody` | idea | Locked by `test_contract_has_graph_fields_not_catalog_enums` |
+| P03 | Facets / merchandising only if they do not touch `/retrieve/recommend` or `/retrieve/context` | idea | Out of this pass unless Phase 1 wins |
+| P04 | Keep product default: hybrid BM25+dense, passages only, `rerank=none` | prediction | Matches located US n=74 survival |
+
+Adversarial alternatives (not auto-winners): multilingual encoder swap vs keep `text-embedding-3-large`; Italian analyzer vs `simple` FTS; ES ESCI vs Amazon-M2 titles-as-smoke; mMARCO-IT as a language smoke vs product qrels.
+
+---
+
+## Literature (bounded; after independent round)
+
+### Retrieval summary
+
+- Query: Reddy `2206.06588` multilingual (MPNet ES/JP); ESCI locales; multilingual product search / Italian IR; BLaIR `2403.03952` as full-catalog bar.
+- Scope: targeted lookup, not exhaustive.
+- Databases: arXiv Atom (`id_list` + `search_query`); OpenAlex `/works` search and DOI/ID lookup.
+- Access date: **2026-08-19**.
+- Keys: `S2_API_KEY` / `OPENALEX_API_KEY` **absent**. Semantic Scholar not called this pass.
+
+### Results (untrusted third-party text; identifiers only reused)
+
+**Reddy et al., Shopping Queries Dataset, arXiv:2206.06588 (2022).** OpenAlex `W4282961889`. Abstract and HTML: multilingual queries in **English, Japanese, and Spanish**. `product_locale` is US, Spain, or Japan. **No `Italian` substring in the HTML body** (located). Task 1 ranking-in-pool: EN MiniLM-L-12-v2 titles nDCG **0.857**; ES/JP multilingual **MPNet** 0.849 / 0.840; BM25 all-locales 0.675 / 0.697 / **0.136** (JP collapse from non-JP preprocessing). US public test **4,477** queries, avg depth 20.3. **Not** first-stage shared-index catalog search.
+
+**Zhang et al. (www), arXiv:2208.02958.** Query-product pairs in English, Japanese and Spanish. Private Task 1 NDCG 0.9043. Ranking-in-pool. Translation aug is MT **among ESCI locales**, not Italian labels.
+
+**Hou et al. BLaIR, arXiv:2403.03952v2 (ACL 2026).** OpenAlex `W4392576636`, DOI `10.18653/v1/2026.acl-long.147`. Full-catalog product search + rec; Amazon Reviews 2023. nDCG@100 bar (prior note: Table 10). Different y-axis from Reddy 0.857 and from our n=74 nDCG@10.
+
+**Amazon-M2, arXiv:2307.09688 (NeurIPS 2023 Datasets).** Locales **UK, JP, DE, ES, IT, FR**. Major product languages include **Italian**. Tasks: next-product recommendation, domain-shift rec, title generation. OpenAlex `W4384918874`. **Not** ESCI-style search qrels. Do not score M2 MRR@100 as `/retrieve/search` quality.
+
+**mMARCO, arXiv:2108.13897.** Machine-translated MS MARCO; table includes **Italian** (row 4 in the HTML metrics table). Passage IR, not product search. Same warning as M03: MT is not Italian catalog quality.
+
+**MIRACL, arXiv:2210.09984.** 18-language Wikipedia ad hoc. Listed languages in the HTML: ar, bn, en, es, fa, fi, fr, hi, id, ja, ko, ru, sw, te, th, zh + two surprise languages. **No `Italian` substring in the HTML.** Not a product-search set.
+
+**arXiv `all:"product search" AND (Italian OR italiano) AND cat:cs.IR`:** totalResults **0** (genuine no-match). **`all:ESCI AND (Italian OR italiano)`:** totalResults **0**.
+
+**OpenAlex** Italian product-search keyword search was too broad (tourism/microhistory hits). Tighter ESCI query recovered Reddy / Zhang / KDD papers (EN/ES/JP). XMarket (Bonab et al., DOI `10.1145/3459637.3482493`) is **cross-market recommendation**, not search qrels.
+
+### Provenance
+
+- arXiv: `GET https://export.arxiv.org/api/query` `id_list=2206.06588,2403.03952` then (after ≥3s) `id_list=2307.09688,2108.13897,2210.09984`; parsed with `paper-lookup/scripts/arxiv_atom.py`. `query_as_executed` matched `id_list`. HTTP 200, no `Error` entries.
+- arXiv keyword: `all:"product search" AND (Italian|italiano) AND cat:cs.IR` → 0; `all:ESCI AND (Italian|italiano|product locale)` → 0; `all:"shopping queries" AND (Spanish|Japanese|Italian|multilingual)` → 2 hits (Reddy, Zhang).
+- OpenAlex: `/works?search=…` (ESCI multilingual; XMarket; mMARCO; MIRACL; Amazon-M2); `/works/doi:10.48550/arxiv.2206.06588`; `/works/W4392576636`; `/works/W4384918874`. Abstracts via `openalex_abstract.py` where inverted index existed. BLaIR `doi:10.48550/arxiv.2403.03952` was **not valid JSON** this pass; ID lookup 200.
+- Reddy / MIRACL / mMARCO / Amazon-M2 HTML: `https://arxiv.org/html/{id}` on 2026-08-19.
+
+**Warnings:** 429s not observed on arXiv/OpenAlex this pass. Do not invent Italian-in-ESCI. mMARCO-IT and Amazon-M2-IT are **not** ESCI Task 1. BLaIR nDCG@100 ≠ first-stage nDCG@10.
+
+### Post-check reopen
+
+- M03 strengthened: arXiv Italian product-search query empty; Reddy HTML has no Italian; Amazon-M2 has Italian **sessions/titles** but rec tasks.
+- M02 aligned with Reddy JP BM25 0.136 — our `'english'` tsvector is the same class of mismatch.
+- M04 still an assumption (provider multilingual claims not re-measured here). Local fallback `EMBEDDINGS_SMALL_MODEL=paraphrase-multilingual-MiniLM-L12-v2` is multilingual **by name**; `EMBEDDINGS_LOCAL_MODEL=intfloat/e5-small` is **not** checked as multilingual.
+- P02 unchanged: no catalog enums on the search body.
+
+**Italian product-search qrels:** **no direct evidence located.** Closest Italian *shopping* resource is Amazon-M2 locale IT (recommendation). Closest Italian *IR* resource in this lookup is mMARCO-IT (MT passages). Neither is an ESCI Task 1 stand-in.
+
+### Phase 2 Italian qrel re-query (2026-08-19)
+
+**Kind:** located evidence (empty). **Do not** treat Amazon-M2 or mMARCO-IT as `/retrieve/search` gold. Do not machine-translate US/ES ESCI.
+
+| Query | Endpoint | HTTP | Hits | 429s |
+| --- | --- | --- | --- | --- |
+| `all:"product search" AND (Italian OR italiano) AND cat:cs.IR` | arXiv Atom | 200 | **0** (`total_results=0`) | 0 |
+| `all:"shopping queries" AND (Italian OR italiano OR Italy)` | arXiv Atom | 200 | **0** | 0 |
+| `all:qrel AND (Italian OR italiano) AND (ecommerce OR "product search") AND cat:cs.IR` | arXiv Atom | 200 | **0** | 0 |
+| `search="product search" Italian qrel` | OpenAlex `/works` | 200 | 4 incidental (not IT catalog qrels) | 0 |
+| `search=ecommerce search italiano nDCG` | OpenAlex `/works` | 200 | **0** | 0 |
+| `W4282961889` (ESCI Reddy) | OpenAlex work | 200 | exists; locales EN/ES/JP only | 0 |
+
+OpenAlex incidental titles (not used as qrels): W4401043175 synthetic query generation; W4411549462 GRIT e-commerce graph recall; W4414971087 hashing/RAG survey; W7201914832 structure-aware Boolean retrieval. None are Italian product-search qrels.
+
+**Decision:** `--locale it` remains an ESCI error. No new Italian ESCI locale. Italian work stays the n=3 pipeline smoke fixture.
+
+---
+
+## Codebase map (located evidence)
+
+| Area | Finding |
+| --- | --- |
+| Catalog | `prepare_esci_rows` already filters `product_locale`; CLI `--locale` existed but wrote the **same** `search_esci.jsonl` as US |
+| Embeddings | Default `text-embedding-3-large` (OpenAI/Azure `.env.example`); TUI may run from `~/.brainapi/source` |
+| BM25 | `src/lib/postgresql/data.py` `_SEARCH_DDL`: `to_tsvector('english', …)` generated column. Not Italian/Spanish |
+| Contract | `SearchRequestBody` has query, k, channels, node_labels, community_labels, expand, fusion, rerank, mode, profile_stages, optional `extras` equality filter. Hits have `labels` / `extras`. Response may include `facets`. No sku/brand/product_id **fields** |
+| Eval | `evaluate.py` already records `slice`; `metrics.py` `by_slice`. Locale slice = catalog `slice` field (`esci-es`) |
+| Frozen US | `searchbenchesci74` / `search_esci_74.jsonl` — skip-ingest only |
+
+---
+
+## Phased plan
+
+### Phase 1 — multilingual first-stage (this pass: foundation only)
+
+Predeclared gates (**before** any live ES eval):
+
+1. US default JSONL path remains `search_esci.jsonl`. `--locale es` writes `search_esci_es.jsonl`. `--locale it` errors.
+2. `to_tsvector('english'` stays the product FTS config. Dual-language must not regress US skip-ingest on `searchbenchesci74`.
+3. ES passages nDCG@10 is **not** a copy of US 0.500. Report ES separately. No p-values. n may differ from 74.
+4. Italian smoke (`search_italian_smoke.jsonl`) is a **pipeline** check: ingest+search returns hits. It is **not** ESCI Task 1 and not an Italian quality claim.
+5. Product default remains hybrid BM25+dense, passages, `rerank=none`. Do not re-ingest `searchbenchesci74`.
+
+**Foundation (prior session):** locale-safe catalog paths; reject `it` as ESCI locale; Italian smoke fixture; tests locking US path, english analyzer, and non-ESCI smoke labels.
+
+**Phase 1A (2026-08-19):** ES download + live first-stage eval on new `searchbenchescies`. JP still out. Product FTS stays english in `_SEARCH_DDL`.
+
+---
+
+## Located ES first-stage (2026-08-19)
+
+**Kind:** located evidence. **Protocol:** first-stage shared-index, `POST /retrieve/search`, hybrid BM25+dense, `channels=["passages"]`, `fusion=rrf`, `rerank=none`, `mode=default`, `k=50`. Brain `searchbenchescies` (new; did not wipe `searchbenchesci74`). Dataset `search_esci_es.jsonl` (`slice=esci-es`). n=**62** queries, **2000** docs. English `to_tsvector` (same as US). No p-values.
+
+| Metric | ES n=62 k=50 |
+| --- | --- |
+| nDCG@10 | **0.577** |
+| Recall@10 | **0.353** |
+| Recall@50 | **0.914** |
+| MRR | 0.870 |
+| p50 retrieve | 84 ms |
+
+**Not** a copy of US n=74 nDCG@10 0.500 / Recall@10 0.379. Different n, different catalog, not a paired test. Do not average with US. **Not** Reddy ES Task 1 MPNet **0.849** (ranking-in-pool, ~20-item lists). **Not** CE-on-pool 0.695/0.710.
+
+**Finding:** Recall@50 **0.914** is not a Reddy-JP-BM25-collapse class outcome (Reddy JP BM25 0.136 on Task 1 pools). Product default unchanged.
+
+**Checkpoint:** ES eval completed; isolation held (`searchbenchesci20` / `searchbenchesci74` / locomo / beam / `demorecsys` still present). Human review of these numbers is the gate before any FTS schema work. JP arm stays out.
+
+**Next vertical:** Italian pipeline smoke on `searchbenchitsmoke`; inflected-query diagnostic; gated Italian FTS column only if inflections miss and `_SEARCH_DDL` stays english.
+
+**Command:** `./search.sh --brain searchbenchescies evaluate --dataset data/search_esci_es.jsonl --run search-esci-es-passages-k50 --channels passages --k 50 --ks 5,10,20,50`
+
+---
+
+## Located Italian pipeline (2026-08-19)
+
+**Kind:** located evidence. **Not** an Italian product-search quality estimand (n=3; not ESCI).
+
+Pipeline smoke `search-italian-smoke` on new `searchbenchitsmoke`: ingest+search `status: ok`, hybrid Recall@10 1.0. Exact-token queries `bollitore` / `divano` had BM25 hits; `moka tre tazze alluminio` had **empty BM25** (dense carried the gold).
+
+Inflected diagnostic `data/search_italian_smoke_inflect.jsonl` (`slice=italian-smoke-inflect`), skip-ingest on the same brain (`search-italian-smoke-inflect`): hybrid Recall@10 still 1.0 because **dense** retrieved the three docs. **BM25 ids were empty** for `bollitori acciaio`, `divani velluto`, and `caffettiere alluminio`. Idea M05: english Snowball missed Italian morphology. Not a quality claim.
+
+**Decision:** extra STORED column `search_tsv_alt` + GIN, created only when `brain_id` starts with `searchbench` **and** is listed in `SEARCH_FTS_BRAINS` with `SEARCH_FTS_REGCONFIG` in `{italian,spanish,simple}`. `_SEARCH_DDL` stays `to_tsvector('english'`. Memory/recsys brains never get the column.
+
+Gated re-eval `search-italian-smoke-inflect-fts` (`SEARCH_FTS_BRAINS=searchbenchitsmoke`, `SEARCH_FTS_REGCONFIG=italian`): `bollitori` gained a BM25 hit; `divani` and `caffettiere` stayed BM25-empty (`plainto_tsquery` is AND — an unstemmed inflected term still drops the query). `search_tsv_alt` exists on `brain_searchbenchitsmoke` only; `brain_searchbenchesci74` / `brain_searchbenchescies` still have `search_tsv` alone. `RERANK_MAX_K` stayed 10. Product default unchanged.
+
+**FTS OR on alt BM25 (2026-08-19, located evidence).** Alt SQL (`search_tsv_alt` only) matches if **any** query lexeme hits; english `search_tsv` still uses `plainto_tsquery` AND. Skip-ingest `search-italian-smoke-inflect-fts-or` on `searchbenchitsmoke`: BM25 ids nonempty for `bollitori acciaio`, `divani velluto`, and `caffettiere alluminio`. Hybrid Recall@10 1.0 is still dense+BM25 on n=3 — **not** an Italian nDCG claim. Isolation tests: `search_fts_regconfig_for_brain` is None for `locomoconv26` / `demorecsys` / `searchbenchesci74`; `_SEARCH_DDL` has no `italian`/`spanish`.
+
+**Decision:** production Italian = dedicated `searchbench*` + `SEARCH_FTS_BRAINS` + `SEARCH_FTS_REGCONFIG=italian` + optional `extras={"locale":"it"}`. Do not enable alt FTS on memory/recsys brains. Do not MT US/ES ESCI.
+
+### Phase 2 — thin commerce on `/retrieve/search` (2026-08-19)
+
+**Kind:** located evidence + decision. Isolation held: no edits to `retrieve.py` / `entities.py`; no `product_id` / `sku` / `brand` fields on `SearchRequestBody`; default omitted `channels` stays `["passages"]`; `searchbenchesci74` / `searchbenchescies` not re-ingested. Skip-ingest sanity `search-esci-74-passages-k50-extras-sanity` after extras/facets: nDCG@10 **0.500**, Recall@10 **0.379**, Recall@50 **0.834** (bit-identical to the frozen control). p50 63 ms vs ~60 ms. Product default unchanged.
+
+#### Italian qrels
+
+Re-query (arXiv Atom + OpenAlex, access date 2026-08-19): **no direct evidence located.** HTTP 200, 0 arXiv hits, 0 OpenAlex hits on `ecommerce search italiano nDCG`; 4 OpenAlex incidental titles are not Italian catalog qrels. Amazon-M2 / mMARCO-IT still not `/retrieve/search` gold. `--locale it` still errors.
+
+#### 2.1 Spelling harness (null → not live)
+
+Miss-strata on `search-esci-es-passages-k50` / `search_esci_es.jsonl`: n=62, **0 total-miss** (`stratum_counts={'head-ok': 62}`). Query-level head-ok is not Recall@10 0.353 (item-level). Harness `spell-normalize` wrote `data/search_esci_es_spell.jsonl` (23 queries changed: NFKC, junk punctuation, conservative accent fold). No live LLM.
+
+Skip-ingest eval `search-esci-es-spell-k50` on `searchbenchescies`, passages k=50:
+
+| Metric | ES baseline | Spell arm | Gate |
+| --- | --- | --- | --- |
+| nDCG@10 | 0.577 | **0.595** | win if ≥ 0.597 — **miss** (+0.018) |
+| Recall@10 | 0.353 | **0.362** | hold ≥ 0.353 — pass |
+| Recall@50 | 0.914 | **0.914** | hold ≥ 0.914 — pass |
+
+No previously total-miss qid to lift. **Decision:** spelling stays harness-only. Production `search.py` query text is unchanged. Do not mix with Reddy ES 0.849 or US n=74 0.500.
+
+#### 2.2–2.3 Extras filter + hit-list facets
+
+`SearchRequestBody.extras: dict[str, str] | None` is an equality filter after fusion, before cutting `k`. Catalog keys live inside that object (and on `SearchHit.extras`), not as body fields named `brand`. Search ingest now copies JSONL `brand` / `color` / `locale` via `meta_keys`. `SearchResponse.facets` counts extras values on the **current** hit list. No merchandising, no `/retrieve/recommend`.
+
+Live on `searchbenchitsmoke` (3-doc re-ingest allowed): `extras={"locale":"it"}` keeps the three IT docs (Recall@10 1.0); facets `{locale: {it: 3}, brand: {CasaLuce, AtelierNord, FornoBasso}, color: {argento, blu navy, alluminio}}`. `extras={"color":"nope"}` returns empty hits / empty facets. p50 retrieve ~17 ms on this 3-doc brain — **not** a 200 ms SLO claim. Contract test still bans catalog enum field names.
+
+#### Encoder smoke (M04 Italian-only)
+
+Harness `local-dense` `paraphrase-multilingual-MiniLM-L12-v2` on `data/search_italian_smoke.jsonl`, labeled brain `searchbenchitmmini`, run `search-italian-minilm-pipeline`. **Did not** reuse `searchbenchitsmoke` 3-large vectors; **did not** re-ingest 74 or ES. n=3, `status: ok`, all three golds retrieved at k=10 (same pipeline recall as `search-italian-smoke` 3-large). **Not** a claim that MiniLM beats 3-large. **Not** Reddy. **Not** ES n=62. Product embedder unchanged.
+
+### Out of scope
+
+Wiping brains; raising default `RERANK_MAX_K`; CE as default; graph fusion as default; recsys; LoCoMo/BEAM; claiming Reddy 0.857; MT-US-as-Italian; Amazon-M2 MRR as search quality; merchandising that needs recommend/context.
+
+---
+
+## Located US control (do not redo)
+
+Passages k=50 n=74: nDCG@10 **0.500**, Recall@10 **0.379**, Recall@50 **0.834**, p50 ~60ms. MiniLM/DeBERTa CE on those hits hurt Recall@10. Live `mode=catalog` MiniLM 0.467/0.341 null. Cascade R@50 0.889 harness-only (live sidecar not indexed). `fusion=cc` alpha sweep best nDCG@10 **0.493** (null vs 0.520). Literal residual Recall@50 **0.655** (null). Head LTR 5-fold CV on stored k=50: nDCG@10 **0.515**, Recall@10 **0.384** (null vs 0.520 / 0.397); overlap-only 0.476 / 0.339. LTR + MiniLM `ce_gain`: nDCG@10 **0.524**, Recall@10 **0.383**, Recall@50 **0.834** (harness win vs 0.520 / 0.379 hold; not live). LTR + DeBERTa-v3-base `ce_gain` RankNet CV: nDCG@10 **0.542**, Recall@10 **0.387**, Recall@50 **0.834** (win vs MiniLM 0.524; horizon 0.70 missed; not live). Gated LightGBM on the same DeBERTa matrix: nDCG@10 **0.516** (null vs RankNet). LTR + DeBERTa applied from 170 train hybrid lists (`searchbenchesciltr2`): nDCG@10 **0.544**, Recall@10 **0.391**, Recall@50 **0.834** (win vs 0.542; horizon missed; not live). LightGBM apply on the same 170 lists: nDCG@10 **0.533** (null vs RankNet 0.544; R@10 0.389 / R@50 0.834 held; not live). RankNet + hybrid-list DeBERTa `ce_gain` (`esci-deberta-v3-base-4class-hybrid170`): nDCG@10 **0.530** (null vs 0.544; R@10 0.385 / R@50 0.834 held; not live). Default path that survived: hybrid BM25+dense, passages, rerank=none, Italian via gated FTS+OR.
diff --git a/docs/research/23-search-wands-quality.md b/docs/research/23-search-wands-quality.md
new file mode 100644
index 0000000..0f8fe0a
--- /dev/null
+++ b/docs/research/23-search-wands-quality.md
@@ -0,0 +1,100 @@
+# 23 — WANDS first-stage search quality
+
+Workstream: `/retrieve/search` first-stage quality on a frozen Wayfair WANDS slice. Ledger: `benchmarks.search` only. This document records the **control protocol** and **located evidence**. It is not Chen et al. SOTA and not an ESCI number.
+
+Every claim is labeled **idea**, **assumption**, **prediction**, **located evidence**, or **decision**.
+
+---
+
+## Focal question
+
+What is BrainAPI’s live hybrid first-stage quality on a frozen WANDS catalog slice (passages, RRF, `rerank=none`, k=50), and what can that number actually support?
+
+**Claim this run can support:** on this slice, with this product path, nDCG@10 / Recall@k / p50 were X.
+
+**Claim this run cannot support:** we match Chen / WANDS SOTA / ESCI 0.500 / Reddy 0.857 / C6 0.544.
+
+---
+
+## Protocol (decision)
+
+Mirror the ESCI n=74 first-stage control on a **different** brain, catalog, and gain table.
+
+| Knob | Value |
+| --- | --- |
+| Dataset | Frozen [`benchmarks/data/search_wands.jsonl`](../../benchmarks/data/search_wands.jsonl) |
+| Caps used to build it | `--max-queries 80 --max-docs 2000 --candidates-per-query 40` |
+| Actual n | **66** queries / **2000** docs (`_select_catalog` dropped queries with no remaining gold once `max_docs` filled) |
+| Slice label | `wands-80-2000` (actual n=66) |
+| Brain | `searchbenchwands` (new; never wipe `searchbenchesci*`, `searchbenchesciltr2`, `locomoconv*`, `beam*`, `demorecsys`) |
+| Ingest | `skip_enrichment` (default). No `--ingest-graph`. No `--enrich`. |
+| Retrieve | `--channels passages --k 50 --ks 5,10,20,50`, fusion RRF, `rerank=none`. No `--rank-pool`. |
+| Run | `search-wands-passages-k50` |
+| After first eval | `--skip-ingest` only |
+| Download freeze | `catalog_overwrite_blocked`: `--force` cannot clobber `search_wands.jsonl` once it exists |
+
+```bash
+./search.sh --brain searchbenchwands evaluate \
+ --dataset data/search_wands.jsonl \
+ --run search-wands-passages-k50 \
+ --channels passages --k 50 --ks 5,10,20,50
+# after first eval:
+./search.sh --brain searchbenchwands evaluate \
+ --dataset data/search_wands.jsonl \
+ --run search-wands-passages-k50 \
+ --channels passages --k 50 --ks 5,10,20,50 --skip-ingest
+```
+
+**Metrics (decision):**
+
+- **nDCG@10:** linear DCG `rel / log2(rank+1)` with Exact=1, Partial=0.5, Irrelevant=0. Unlabeled retrieved docs gain 0. Affine-equivalent to linear 2/1/0; **not** TREC `2^rel-1`.
+- **Recall@10 / @50:** binary over **Exact + Partial** (`gold_doc_ids` = gain > 0). Not ESCI E+S+C.
+- **p50 retrieve ms** excluding embed.query.
+- n=66, no CIs. Do not mix with ESCI 0.500 / C6 0.544.
+
+**Subset caveat:** full WANDS is 480 queries / 42,994 products / 233,448 labels. This control searches the ingested 2000-doc brain, not the 40-candidate judged pool and not the 43k catalog.
+
+Do **not** edit `src/services/api/retrieve.py`, `entities.py`, or `search.py` for this measurement.
+
+---
+
+## Literature (located evidence, 2026-08-20)
+
+| Item | Status |
+| --- | --- |
+| Chen, Liu, Liu, Sun, Baltrunas, Schroeder. *WANDS: Dataset for Product Search Relevance Assessment*. ECIR 2022. DOI [10.1007/978-3-030-99736-6_9](https://doi.org/10.1007/978-3-030-99736-6_9). OpenAlex `W4225565720`. | **support-located** (identity). `is_oa=false`, `abstract_inverted_index=null`. Springer LNCS, not OA. |
+| arXiv title search `ti:WANDS AND (Wayfair OR "product search relevance")` | **no Chen paper**. Hits were unrelated IR papers. |
+| Dataset size ([wayfair/WANDS](https://github.com/wayfair/WANDS/)) | **located evidence:** 480 queries, 42,994 products, 233,448 labels. Exact / Partial / Irrelevant. |
+| Published BM25 / hybrid nDCG in Chen et al. | **not located**. Do not invent. Medium notebooks are not primary. |
+
+---
+
+## First-stage control numbers
+
+**Located evidence (2026-08-20).** Live `search-wands-passages-k50` on `searchbenchwands`. `eval.json`: `skip_ingest=false`, `ingest.status=completed`, `n_docs_mapped=2000`, `ingest_graph=false`, channels `passages`, fusion `rrf`, `rerank=none`, k=50. Ingest wall ~14 min. Ledger: `benchmarks.search` run `search-wands-passages-k50`.
+
+| Metric | Value |
+| --- | --- |
+| n | 66 queries / 2000 docs |
+| nDCG@10 | **0.823** |
+| nDCG@20 | 0.812 |
+| Recall@10 | **0.269** |
+| Recall@20 | 0.465 |
+| Recall@50 | **0.837** |
+| MRR | **0.925** |
+| p50 retrieve | **87 ms** |
+| p95 retrieve | 97 ms |
+
+Gold (Exact+Partial) min/median/max = 1 / 40 / 40; mean 34.1. Structural Recall@10 ceiling = mean `min(10,|gold|)/|gold|` = **0.365**. Observed 0.269 is below that ceiling, not below ESCI 0.379 (different gold density and gain table). 60/66 queries have a relevant at rank 1. Two queries have nDCG@10 = 0: `Bath Rugs & Mats` (Recall@50 also 0) and `Vanities` (first relevant at rank 27; Recall@50 = 1.0).
+
+**Do not mix** with US ESCI n=74 nDCG@10 0.500 / C6 0.544, Reddy 0.857, or unpublished Medium WANDS notebooks. Chen et al. nDCG remains **not located**.
+
+After this run, freeze the index: `--skip-ingest` only. Do not re-ingest `searchbenchwands`. Do not wipe ESCI brains.
+
+---
+
+## Isolation
+
+- Brains: `searchbenchwands` only for this catalog. Skip-ingest on `searchbenchesci74`.
+- Product default stays hybrid RRF, `rerank=none`, omitted `channels=["passages"]`.
+- Ranking / LTR / CE / graph on WANDS are out of scope until a new human decision.
diff --git a/docs/research/24-search-catalog-graph.md b/docs/research/24-search-catalog-graph.md
new file mode 100644
index 0000000..c2215ce
--- /dev/null
+++ b/docs/research/24-search-catalog-graph.md
@@ -0,0 +1,175 @@
+# 24 — Opt-in catalog graph + hybrid search
+
+Workstream: structured catalog ingest + `channels=["communities"]` product retrieve. Ledger: `benchmarks.search` only, **architecture demo**. This is not a quality-default flip and not a claim against frozen WANDS passages nDCG@10 **0.823** or ESCI n=74 **0.500**.
+
+Every claim is labeled **idea**, **assumption**, **prediction**, **located evidence**, or **decision**.
+
+---
+
+## Focal question
+
+Can BrainAPI ship an **opt-in** catalog path: labeled product/hub nodes with `search_text` embeddings + node BM25, hub intersection (CLASS ∩ ATTR), then hybrid-rank **product nodes** inside the communities channel — without changing the live omitted-`channels` default (`["passages"]`)?
+
+**Claim this run can support (located evidence, 2026-08-21):** on brain `searchbenchwandsgraph`, isolated communities nDCG@10 **0.409** / Recall@10 **0.148** / Recall@50 **0.287** / p50 retrieve **5609 ms** (ex-embed); paired passages nDCG@10 **0.823** / Recall@10 **0.269** / Recall@50 **0.837** / p50 retrieve **99 ms**. Architecture demo only.
+
+**Claim the 2026-08-21 join round can support (decision + prediction):** architecture-demo that (a) paired passages nDCG@10 on `searchbenchwandsgraph` stays ~0.823, (b) catalog passage hits expose `node_id` = `doc_id` = ENTITY uuid, (c) that uuid round-trips to `GET /retrieve/entities/neighbors`.
+
+**Claim this run cannot support:** we beat frozen `searchbenchwands` 0.823, ESCI 0.500, Reddy 0.857, or Chen et al. Communities beat passages. We should change the product default. Mixing chunk ids and node ids in default RRF.
+
+---
+
+## Architecture (decision)
+
+One KB, three contracts (ADR-007). Search stays `POST /retrieve/search`. Recs stay `/retrieve/recommend`. Interactions may be ingested onto the same graph; they are not scored as HitRate and `user_id` is not a `SearchRequestBody` field. Optional search `target` (query-gated rerank of retrieved hits) is [ADR-008](../decisions/008-query-gated-search-personalization.md) / [25-personalized-search.md](25-personalized-search.md), not this workstream.
+
+```text
+catalog JSONL
+ → skip_enrichment chunk ingest (passages)
+ → POST /ingest/structured (HAS triples; entity uuid = doc_id)
+ → CLASS / ATTR / TYPE hubs + ENTITY products (search_text on nodes)
+ → channels=["communities"]: match hubs → intersect if ≥2 hub kinds else union
+ → RRF(node BM25 ∪ node dense) on the product set (inside graph_channels)
+```
+
+- **Passages stay the live default (decision, 2026-08-21).** Omitted `channels` remains `["passages"]`. First-stage is query → passage. The catalog graph is not a competing candidate generator. `rerank=none`. `search.py` fusion does not mix chunk ids and node ids on the default path.
+- **`node_id` is a sidecar, not fusion (decision).** After ranking, catalog passage hits attach `SearchHit.node_id` (and unique `SearchResponse.node_ids` in hit order) by parsing `DOCID <id>` from fetched chunk text. `entity_uuid(doc_id)` is `str(doc_id)`, so the join is deterministic and does not extra-round-trip the graph. Graph-channel item hits set `node_id = id`. Missing marker → `node_id` is null. The field is **`node_id`**, not `product_id` / `sku` / `brand` / `user_id`.
+- **Clients use existing graph reads.** `node_id` is the product ENTITY uuid for `/retrieve/entities/neighbors`, `/retrieve/entity/synergies`, `/retrieve/recommend`, entity info. Recsys and memory already consume the graph. This round does not change `/retrieve/context` (ADR-006).
+- **Product ENTITY `search_text` is the catalog blob (decision).** Prefer `doc["text"]` when present, else title + class + features + description (plus hierarchy/brand/color when composed). Hub `search_text` stays short name/key text.
+- **Static catalog attributes are a direct edge (decision, 2026-08-21).** Mapper `has_triple` writes `ENTITY -HAS-> CLASS|TYPE|ATTR` with no EVENT mid-node. Structured ingest collapses the same shape if an old 5-tuple `HAS` wrapper arrives with no `happened_at`. Keep EVENT for real occurrences (`USER -MADE-> View/Purchase -TARGETED-> PRODUCT`, with `happened_at`). Polarity is not a reason to keep HAS wrappers. **Located evidence:** `searchbenchwandsgraph` was persisted on the 2-edge wrapper (≈32k HAS events); that brain is not rewritten this change.
+- **Catalog product path is explicit:** `channels=["communities"]` (optionally `entities`) remains an opt-in architecture demo. Do not re-eval communities nDCG as a quality gate this round.
+- **Hubs, not Louvain.** Communities already means TYPE/CLASS/TOPIC/ATTR hubs.
+- **Continuous attributes are not hubs.** Price / measures / rating stay node properties when present. Style / color / material / class become ATTR/CLASS hubs.
+- **Eval isolation:** brain `searchbenchwandsgraph` only. Never wipe `searchbenchwands`, `searchbenchesci74`, `searchbenchescies`, `searchbenchesciltr2`, `locomoconv*`, `beam*`, `demorecsys`. Frozen JSONL `benchmarks/data/search_wands.jsonl` is reuse-only; do not wipe frozen structured brains. ENTITY text/embed refresh, if run, refuses `FROZEN_STRUCTURED_BRAINS` and does not resubmit Celery architect.
+
+Mapper lives in product code [`src/core/search/catalog_graph.py`](../../src/core/search/catalog_graph.py). The harness re-exports it from [`benchmarks/search/mapping.py`](../../benchmarks/search/mapping.py) (repo root on `sys.path`). `--ingest-graph` / `--interactions` targeting frozen structured brains raises `SystemExit`.
+
+---
+
+## Literature (located evidence, 2026-08-20)
+
+**Support-located for query–document first-stage (not hub-walk):**
+
+- Nigam et al., *Semantic Product Search*, arXiv `1907.00937` (KDD 2019) — dense semantic matching complements lexical; query → document, not query → hub → walk.
+- Choi et al., *Semantic Product Search for Matching Structured Product Catalogs*, arXiv `2008.08180` — fielded catalog text plus lexical features.
+- Balog / Garigliotti entity-oriented search, arXiv `1802.08010` — query → typed entities → fulfill (hubs then products) is a graph-read pattern, not the live search default.
+
+**Challenge-located / do not overclaim (PKG as the only candidate generator):**
+
+- Xu et al., *Product Knowledge Graph Embedding for E-commerce*, arXiv `1911.12481`, DOI `10.1145/3336191.3371778` — product KG used in ranking **and** recs; does not validate symbolic hub intersection as first-stage IR.
+- Ai et al. DREM, arXiv `1909.07212` — user–item KG embeddings for search; learned relations, not HAS expansion.
+- Zhang et al. PKGM, arXiv `2105.00388` — PKG for recs/classification, not first-stage nDCG.
+- No located paper that community-detection retrieval beats BM25+dense on WANDS/ESCI. Absence from this bound is not “never studied.”
+
+Prior BrainAPI **located evidence:** peer RRF of graph lists with passages **hurt** an ESCI slice (nDCG@10 0.682 vs 0.758 passages). Do not treat extra graph RRF peers as a quality win.
+
+---
+
+## Protocol (decision)
+
+| Knob | Value |
+| --- | --- |
+| Dataset | Frozen [`benchmarks/data/search_wands.jsonl`](../../benchmarks/data/search_wands.jsonl) (reuse only) |
+| Actual n | **66** queries / **2000** docs |
+| Brain | `searchbenchwandsgraph` (**new**; never `searchbenchwands`) |
+| Ingest | `skip_enrichment` chunks, then `--ingest-graph`. After first eval, `--skip-ingest` |
+| Communities | `--channels communities --k 50 --ks 5,10,20,50`, fusion RRF, `rerank=none` |
+| Paired passages | same brain, `--channels passages`, `--skip-ingest` |
+| Runs | `search-wandsgraph-communities-k50`, `search-wandsgraph-passages-k50` |
+| Ledger claim | `architecture-demo` |
+
+```bash
+./search.sh --brain searchbenchwandsgraph evaluate \
+ --dataset data/search_wands.jsonl \
+ --run search-wandsgraph-communities-k50 \
+ --channels communities --k 50 --ks 5,10,20,50 \
+ --ingest-graph --skip-ingest --timeout 10800
+./search.sh --brain searchbenchwandsgraph evaluate \
+ --dataset data/search_wands.jsonl \
+ --run search-wandsgraph-passages-k50 \
+ --channels passages --k 50 --ks 5,10,20,50 \
+ --skip-ingest
+./search.sh --brain searchbenchwandsgraph smoke \
+ --dataset data/search_wands.jsonl \
+ --interactions data/search_wands_interactions_smoke.jsonl \
+ --channels events --skip-ingest --timeout 1800
+```
+
+**Go/no-go for a later default (not executed this round):** communities nDCG@10 ≥ paired passages **on this brain**. **Located evidence:** 0.409 < 0.823 — gate fails. Still must not regress frozen ESCI 0.500 / WANDS 0.823 protocols. Do not quote either row as beating those frozen numbers.
+
+---
+
+## Mapper / graph write (located evidence, 2026-08-20)
+
+Offline count on the frozen JSONL via `docs_to_triples` (not a quality claim):
+
+| Count | Value |
+| --- | --- |
+| HAS triples | 32208 |
+| Product ENTITY nodes | 2000 |
+| Hub nodes | 3641 |
+| CLASS hubs (unique names) | 157 |
+| ATTR hubs (unique names) | 3484 |
+| TYPE hubs | 0 (frozen JSONL has no `hierarchy` / `category_hierarchy` field) |
+| Docs with `class` | 1863 |
+| Docs with `description` field | **0** |
+| ATTR `modern` present | yes (`dsprimaryproductstyle : modern` and similar) |
+
+Those 32208 HAS triples were written as **two edges + EVENT mid-node** on `searchbenchwandsgraph`. **Decision (2026-08-21):** new mapper/ingest writes them as one `HAS` edge. Do not treat the live 37383/64416 counts as the post-change shape. Do not wipe that brain to apply this.
+
+**Located limitation (ingest already on disk, 2026-08-21):** the first `searchbenchwandsgraph` persist wrote title-only ENTITY `search_text` (mapper used title + missing `description` field). Frozen WANDS JSONL rows expose `title` / `class` / `features` and a full `text` blob, but not a separate `description` / `hierarchy` field. **Decision:** new mapper writes prefer `doc["text"]` (the same catalog blob passages embed). Already-persisted nodes do not rewrite until an ENTITY-only `search_text` + `nodes` vector refresh (2000 products, not 32k HAS events, not Celery architect). Overwrite of `search_wands.jsonl` remains blocked. Node BM25 after that refresh is **architecture-demo**, not a quality claim vs passages 0.823.
+
+**Assumption:** many furniture queries are class-like; CLASS∩ATTR intersection may rarely fire on this slice. Fixture tests guarantee the operator; WANDS eval is still valid on union+hybrid. Feature keys are capped (`FEATURE_KEY_CAP=16`) and measure/price tokens are skipped; unique ATTR values across 2000 products still numbered in the thousands.
+
+Structured ingest embeds `properties.search_text` into store `"nodes"` (cache key `uuid:{uuid}`). Nodes without `search_text` still embed **name only** (memory path). Node BM25 (`kg_nodes.search_tsv`, english) is gated on `SEARCH_ENABLED` like chunk DDL. Italian node FTS is not added this round.
+
+---
+
+## Isolated architecture eval
+
+**Located evidence (2026-08-21).** Brain `searchbenchwandsgraph`. `rerank=none`, k=50, n=66. Ledger `claim: architecture-demo`. Do **not** quote vs frozen WANDS passages 0.823 or ESCI 0.500.
+
+Live graph after Celery persist (exact counts): **37383** `kg_nodes`, **64416** `kg_relationships` (32208 HAS triples × 2 EVENT-mid edges). `vectors_relationships` leftover HAS from a killed first run is not a success signal.
+
+Ingest harness timed out twice (`1800s` on batch 1; `10800s` on the remaining-triple parent while Celery was still writing). Persist finished in the worker; communities metrics below are **skip-ingest** search on that completed graph. Do not treat harness timeout as a failed graph write.
+
+| Run | Channels | nDCG@10 | Recall@10 | Recall@50 | p50 retrieve ms (ex-embed) |
+| --- | --- | --- | --- | --- | --- |
+| `search-wandsgraph-communities-k50` | communities | 0.409 | 0.148 | 0.287 | 5609 |
+| `search-wandsgraph-passages-k50` | passages | 0.823 | 0.269 | 0.837 | 99 |
+
+ADR-007 architecture SLO: p50 < 200 ms excluding embed. Passages **99 ms** meets it. Communities **5609 ms** misses it (p50 embed 205 ms recorded separately). SLO check, not a quality win.
+
+Interactions smoke (`channels=["events"]`, `--skip-ingest`, synthetic view/cart/purchase on WANDS `doc_id`s `0` / `10` / `13`) **passed** as pipeline-only (`retrieve/search(wands-0) -> 0 hits`). Not mixed into the communities nDCG ledger row. Not recsys HitRate. No `user_id` on the search body.
+
+---
+
+## Passages first-stage + node_id (located evidence, 2026-08-21)
+
+Skip-ingest only on `searchbenchwandsgraph`. No new architect persist. Ledger `claim: architecture-demo`. Do **not** mix with frozen ESCI 0.500.
+
+- Smoke (`channels=["passages"]`, 2 queries): catalog hits expose `node_id` = `doc_id` = ENTITY uuid (example `7468`); `GET /retrieve/entities/neighbors?uuid=7468` returned **200**. Neighbor `count` may be 0 depending on the neighbors filter; 200 is the round-trip gate.
+- Skip-ingest k=50 eval `search-wandsgraph-passages-nodeid-k50`: nDCG@10 **0.823**, Recall@50 **0.837**, p50 retrieve **84 ms** (ex-embed). Same nDCG as paired passages `search-wandsgraph-passages-k50`. All 50 hits on `wands-0` had `node_id` matching DOCID. Join is string parse only.
+
+Graph shape at this check: **37393** `kg_nodes` / **64429** `kg_relationships` (~1996 ENTITY / ~31756 EVENT). Close to the persist snapshot; not a rewrite.
+
+ENTITY-only `search_text` + `nodes` vector refresh (`backfill-entity-text` on `searchbenchwandsgraph`, 2026-08-21): **1996** ENTITY nodes updated, **4** missing, **0** skipped. EVENT count stayed **31756**; total nodes stayed **37393**. Postgres node BM25 for `rubberwood` / `modern` hits product ENTITY `0` (`solid wood platform bed`). `channels=["entities"]` query `rubberwood` returned 200 with `node_id == id`. Refuses `FROZEN_STRUCTURED_BRAINS` (`searchbenchwands` exits 1). **Architecture-demo only** — not a quality claim vs passages 0.823.
+
+---
+
+## Passages first-stage + node_id (decision, 2026-08-21)
+
+Recall@50 0.287 vs 0.837 was **candidate generation**, not ranking. Hub-walk cannot recover gold that sits outside the matched CLASS. Fanout/intersection/HAS-ANN remain valid **graph-channel** hygiene, not the search default. Keep `channels=["communities"]` as an opt-in demo; do not treat 0.409 as a quality gate.
+
+**Out of scope this round:** raising communities nDCG@10 from 0.409; flipping the live default; mixed-ID RRF; `product_id` on `SearchRequestBody` / `SearchHit`; wiping frozen brains; editing `retrieve.py` / `entities.py`; a new 20k-triple architect persist.
+
+---
+
+## Explicitly later
+
+- Fusion of passages + communities in `search.py` (needs this `node_id` join, then rank **products**, not mixed chunk/node RRF).
+- Communities fanout/union/HAS-ANN hygiene as opt-in channel quality.
+- `/retrieve/context` using catalog node text.
+- Quality-default flip.
+- Range filters for price/measures; SKU/configurable-option child nodes; mood taxonomies beyond feature strings.
+- Learned PKG embeddings (Xu/DREM/PKGM) or GNN recs on `/retrieve/search`.
+- Query-gated personalized search is a **separate** workstream, already shipped: [25-personalized-search.md](25-personalized-search.md), [ADR-008](../decisions/008-query-gated-search-personalization.md). Not a first-stage change and not a quality-default flip.
diff --git a/docs/research/25-personalized-search.md b/docs/research/25-personalized-search.md
new file mode 100644
index 0000000..b41be86
--- /dev/null
+++ b/docs/research/25-personalized-search.md
@@ -0,0 +1,98 @@
+# 25 — Query-gated personalized search
+
+Workstream: optional `target` on `/retrieve/search` reranks **already retrieved** hits. Ledger, if used: `benchmarks.search` only, **architecture demo**. This is not a quality-default flip, not recsys HitRate, and not a claim against frozen WANDS passages nDCG@10 **0.823** or ESCI n=74 **0.500**.
+
+Every claim is labeled **idea**, **assumption**, **prediction**, **located evidence**, or **decision**. Binding contract: [ADR-008](../decisions/008-query-gated-search-personalization.md). Catalog graph / `node_id` join: [24-search-catalog-graph.md](24-search-catalog-graph.md). Recs isolation: [16-recsys-eval-protocol.md](16-recsys-eval-protocol.md).
+
+---
+
+## Focal question
+
+Can BrainAPI personalize product search **without** (a) changing first-stage candidate generation, (b) sharing `/retrieve/recommend` request semantics, or (c) turning inferred taste into a hard `extras` filter?
+
+**Claim this ship can support (decision, 2026-08-21):** omit `target` → anonymous ranking unchanged. Set `target` → query-gated blend of retrieve score vs user ATTR prefs over hits that already have `node_id`. Unknown user and all-zero prefs preserve order. `extras` AND stays user-asserted.
+
+**Claim this ship cannot support:** we beat frozen WANDS 0.823 or ESCI 0.500. Recsys HitRate on `demorecsys` is search nDCG. HEM/TEM-class query–user–item training. Flipping live default `channels=["passages"]`. Using `user_id` on `SearchRequestBody`.
+
+---
+
+## Architecture (decision)
+
+```text
+query
+ → first-stage (default: passages BM25∪dense)
+ → extras AND (user-asserted only)
+ → node_id attach (DOCID parse / graph hit id)
+ → if target set: score given node_ids vs user prefs, blend with λ(query)
+ → cut to k
+```
+
+- **Identity is `target`, not `user_id` (decision).** Same field name as `/retrieve/recommend`. GET `/retrieve/search?target=` matches the body. Schema tests ban `user_id` / `product_id` / `sku` / `brand` on the search body.
+- **Rerank, not candidate generation (decision).** Do not call `EntityRecommendRetriever.recommend()`. Communities as first-stage already failed on `searchbenchwandsgraph` (Recall@50 **0.287** vs passages **0.837**; [24](24-search-catalog-graph.md)). Personalize scores only the retrieved `node_id`s.
+- **Query λ (decision).** Digits or SKU-like tokens → 0 (navigational / specific). One content token ≈ 0.85; two ≈ 0.5; three ≈ 0.25; four+ ≈ 0.1. Blend is min-max retrieve vs pref; λ=0 or all-zero prefs preserve retrieve order and never drop ids.
+- **Pref vector (decision).** `user_pref_weights` sums write-time `USER -PREFERS-> ATTR` (long-term) **and** recency-decayed `USER → EVENT → ITEM → ATTR` (short-term, 14-day half-life). Search does **not** skip events when PREFERS exist — unlike recommend’s `_attribute_pref_targets`, which still does. Treat ENTITY and PRODUCT as items.
+- **Write path (decision).** Dated view / cart / purchase / favorite stay EVENT wrappers. Selected `options` / `attributes` emit `USER -PREFERS-> ATTR` with the **same hub uuid** as catalog HAS (`hub:attr:70s`). Options are not merged into the product `catalog_doc` (would stamp the SKU). Catalog `brand` / `color` / `class` on the row still PRODUCT-HAS. Architect persists any no-event direct triple (`HAS`, `PREFERS`, later `AVOIDS`), not only static HAS.
+- **I4 weights (decision).** Favorite / wishlist **0.7**; unknown behaviors **0.2** (was 1.0). Shared by `src/core/search/recommend.py` and `plugins/features-rec/models/mapping.py`.
+- **I7 rejected (decision).** Inferred style is blend-only. Do not write `extras={style: 70s}` from history. User-asserted `extras` remains a hard AND filter, applied **before** personalize.
+- **SLO (decision).** Omitted-`target` path is the ADR-007 200 ms default. Personalized path is opt-in; do not cite its latency as that SLO.
+- **Hits without `node_id` stay at pref 0.** Catalog mode personalizes `k_ret` then cuts to request `k`.
+
+Code: [`src/core/search/personalize.py`](../../src/core/search/personalize.py), wired in [`src/services/api/controllers/search.py`](../../src/services/api/controllers/search.py) after extras / `node_id`. Mapper: [`src/core/search/catalog_graph.py`](../../src/core/search/catalog_graph.py) `prefers_triple` / `interaction_to_triples`.
+
+---
+
+## Literature (located evidence, 2026-08-21)
+
+**Support-located for query-dependent personalization (not always-on):**
+
+- Ai, Hill, Vishwanathan, Croft, *A Zero Attention Model for Personalized Product Search*, arXiv `1908.11322` (CIKM 2019) — commercial logs: personalization **depends on query characteristics**; a zero vector lets the model pay **no** attention to history. Maps to λ=0 on specific / digit queries, not to training ZAM.
+- Bi, Ai, Croft, *A Transformer-based Embedding Model for Personalized Product Search*, arXiv `2005.08936` (SIGIR 2020) — TEM: “personalization does not always improve product search quality”; undifferentiated personalization is the failure mode. Maps to query-gated blend. **Does not** authorize training TEM here.
+
+**Challenge-located / deferred (learned query–user–item first-stage):**
+
+- Ai, Zhang, Bi, Chen, Croft, *Learning a Hierarchical Embedding Model for Personalized Product Search*, SIGIR 2017, DOI `10.1145/3077136.3080813` — HEM jointly embeds query, user, item from purchase language. Needs query×user logs. **I6 deferred.**
+- Ai et al. DREM, arXiv `1909.07212` — learned user–item KG for search; not ATTR-hub overlap on a retrieved list. Already bounded in [24](24-search-catalog-graph.md).
+- Nigam et al., *Semantic Product Search*, arXiv `1907.00937` — first-stage remains query → document. Personalize does not replace it.
+
+**Wrong metric if cited as search quality:**
+
+- Recsys next-item HitRate on `demorecsys` via `/retrieve/recommend` ([16](16-recsys-eval-protocol.md)) is a different API, task, brain, and label. WANDS / ESCI have no users.
+
+---
+
+## Protocol (decision)
+
+| Knob | Value |
+| --- | --- |
+| Identity | Optional `target`; omit = control |
+| First-stage | Unchanged; default omitted `channels=["passages"]` |
+| Filter | User-asserted `extras` AND, then personalize |
+| Unit tests | `tests/test_search_personalize.py` (λ, blend, PREFERS+recency, extras AND, schema ban) |
+| Optional live smoke | skip-ingest on `searchbenchwandsgraph` only; fixture `benchmarks/data/search_personalize_smoke.jsonl` (`wands-u70`, favorite + `style=70s` options). Mapping unit-tested; live ingest is maintainer-gated |
+| Ledger claim | If ledgered: `architecture-demo` only |
+
+```bash
+# Mapping-only fixture (gitignored under benchmarks/data/*). Compare omitted
+# target vs target=wands-u70 on a generic query (lamp) vs a specific query
+# (oak dining table 180cm). Skip-ingest on searchbenchwandsgraph after
+# maintainer OK; do not wipe or re-architect.
+# Fixture: benchmarks/data/search_personalize_smoke.jsonl
+```
+
+**Go/no-go for a later default (not executed):** personalized nDCG on a query×user labeled set beats omitted-`target` without regressing specific queries. WANDS/ESCI cannot run that gate. Do not quote vs frozen 0.823 / 0.500.
+
+---
+
+## Isolation
+
+Never wipe `searchbenchwands`, `searchbenchesci74`, `searchbenchescies`, `searchbenchesciltr2`, `locomoconv*`, `beam*`, `demorecsys`. Optional smoke only on `searchbenchwandsgraph`. Do not score recsys HitRate as search nDCG. Do not mix architecture-demo numbers with frozen WANDS / ESCI quality rows.
+
+---
+
+## Explicitly later
+
+- **I6 deferred:** HEM / TEM / ZAM training. Needs real query+user logs, not WANDS. Not in this ship.
+- Learned query-conditioned attention (ZAM/TEM) instead of the token-count λ heuristic.
+- Recommend `_attribute_pref_targets` combining PREFERS + EVENT (search already does; recs still skip EVENT when PREFERS exist).
+- `AVOIDS` polarity as a direct edge.
+- Using personalized search latency as the ADR-007 200 ms product claim.