Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
{
"schema_version": 1,
"decision_id": "p2-date-grounding-dev",
"plan": "P2",
"title": "P2 date grounding (extraction.date_grounding) dev comparison",
"verdict_type": "quality",
"stage": "dev",
"candidate": {
"gbrain": "/workspace/gbrain@5d5375d43",
"config": {
"extraction.date_grounding": "true"
}
},
"baseline": {
"gbrain": "/workspace/gbrain@5bd9e84978126689b11cf65107c0a9aa0e91e11b",
"config": {}
},
"sources": [
{
"kind": "memory-qa",
"split": "dev",
"embed": "real",
"search_pins": {
"search.mode": "balanced",
"search.reranker.enabled": "false",
"search.autocut": "false"
},
"top_k": 10,
"id_field": "id",
"exclude_when": [
"abstention=true"
],
"facts": "conversation",
"qa": {
"mode": "reader",
"runs": 1,
"sessions": 5,
"context": "facts"
},
"id": "locomo-dev-temporal-facts",
"benchmark": "locomo",
"limit": 150,
"categories": [
"temporal"
],
"min_clusters": 3,
"comparisons": [
{
"id": "facts-unresolved-relative",
"metric": "facts_unresolved_share",
"gate": "superiority",
"direction": "lower",
"min_effect": 0,
"cluster_by": "conversation",
"description": "share of saved facts that still carry a relative time expression (yesterday, last week, 3 days ago) and no absolute date"
},
{
"id": "temporal-qa-from-facts",
"metric": "qa_score",
"gate": "superiority",
"direction": "higher",
"min_effect": 0,
"cluster_by": "conversation",
"description": "temporal questions answered by a fixed reader from the saved facts of the top five sessions"
},
{
"id": "temporal-qa-from-facts-by-question",
"metric": "qa_score",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "id"
},
{
"id": "recall-all-5",
"metric": "recall_all_at_5",
"gate": "noninferiority",
"direction": "higher",
"tolerance": 0.01,
"cluster_by": "conversation",
"description": "guardrail: page retrieval unchanged (date grounding touches extraction only)"
},
{
"id": "facts-count",
"metric": "facts_count",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "conversation"
}
],
"estimate_usd": 8
},
{
"kind": "memory-qa",
"split": "dev",
"embed": "real",
"search_pins": {
"search.mode": "balanced",
"search.reranker.enabled": "false",
"search.autocut": "false"
},
"top_k": 10,
"id_field": "id",
"exclude_when": [
"abstention=true"
],
"facts": "conversation",
"qa": {
"mode": "reader",
"runs": 1,
"sessions": 5,
"context": "facts"
},
"id": "locomo-dev-nontemporal-facts",
"benchmark": "locomo",
"limit": 150,
"categories": [
"single-hop",
"multi-hop",
"open-domain"
],
"min_clusters": 3,
"comparisons": [
{
"id": "nontemporal-qa-from-facts",
"metric": "qa_score",
"gate": "noninferiority",
"direction": "higher",
"tolerance": 0.01,
"cluster_by": "conversation",
"description": "guard: non-temporal questions answered from saved facts do not drop by more than one point"
},
{
"id": "nontemporal-qa-by-question",
"metric": "qa_score",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "id"
},
{
"id": "facts-count",
"metric": "facts_count",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "conversation"
},
{
"id": "recall-all-5",
"metric": "recall_all_at_5",
"gate": "noninferiority",
"direction": "higher",
"tolerance": 0.01,
"cluster_by": "conversation"
}
],
"estimate_usd": 8
},
{
"kind": "memory-qa",
"split": "dev",
"embed": "real",
"search_pins": {
"search.mode": "balanced",
"search.reranker.enabled": "false",
"search.autocut": "false"
},
"top_k": 10,
"id_field": "id",
"exclude_when": [
"abstention=true"
],
"facts": "conversation",
"qa": {
"mode": "reader",
"runs": 1,
"sessions": 5,
"context": "facts"
},
"id": "lme-s-dev-temporal-facts",
"benchmark": "lme-s",
"limit": 12,
"categories": [
"temporal-reasoning"
],
"min_clusters": 10,
"comparisons": [
{
"id": "facts-unresolved-relative",
"metric": "facts_unresolved_share",
"gate": "superiority",
"direction": "lower",
"min_effect": 0,
"cluster_by": "conversation",
"description": "share of saved facts that still carry a relative time expression (yesterday, last week, 3 days ago) and no absolute date"
},
{
"id": "temporal-qa-from-facts",
"metric": "qa_score",
"gate": "superiority",
"direction": "higher",
"min_effect": 0,
"cluster_by": "conversation",
"description": "temporal questions answered by a fixed reader from the saved facts of the top five sessions"
},
{
"id": "temporal-qa-from-facts-by-question",
"metric": "qa_score",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "id"
},
{
"id": "recall-all-5",
"metric": "recall_all_at_5",
"gate": "noninferiority",
"direction": "higher",
"tolerance": 0.01,
"cluster_by": "conversation",
"description": "guardrail: page retrieval unchanged (date grounding touches extraction only)"
},
{
"id": "facts-count",
"metric": "facts_count",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "conversation"
}
],
"estimate_usd": 60
}
],
"alpha": 0.05,
"seed": 42,
"draws": 10000,
"budget_usd": 76,
"created_at": "2026-10-04T20:31:50.737Z",
"notes": "Date grounding of extracted facts. Both arms run the facts lane: each conversation's sessions are imported as conversation pages with ISO session dates, gbrain's conversation-facts extractor runs on them (product default model), and the temporal questions are answered by a fixed reader from the saved facts (fact text and stored date) of the top five retrieved sessions, never the raw sessions. Primary: share of saved facts that keep an unresolved relative time expression (judge-free) and temporal QA from facts. Retrieval is a guardrail. Baseline is the candidate branch base (merge-base with master), so only the P2 diff differs."
}