Garry's Opinionated OpenClaw/Hermes Agent Brain
TypeScript
{
"schema_version": 1,
"decision_id": "p3-feedback-guardrail-dev-v2",
"plan": "P3",
"title": "P3 retrieval feedback guardrail (dev): default-on feedback with no ratings must not change retrieval",
"verdict_type": "quality",
"stage": "dev",
"candidate": {
"gbrain": "/workspace/gbrain@5e5fdcbbe55827998dd22df2e16aa777d522520f",
"config": {}
},
"baseline": {
"gbrain": "/workspace/gbrain@5bd9e84978126689b11cf65107c0a9aa0e91e11b",
"config": {}
},
"sources": [
{
"id": "lme-s-dev",
"kind": "memory-qa",
"benchmark": "lme-s",
"split": "dev",
"embed": "real",
"limit": 100,
"search_pins": {
"search.mode": "balanced",
"search.reranker.enabled": "false",
"search.autocut": "false"
},
"top_k": 10,
"id_field": "id",
"exclude_when": [
"abstention=true"
],
"min_clusters": 10,
"comparisons": [
{
"id": "recall-all-5",
"metric": "recall_all_at_5",
"gate": "noninferiority",
"direction": "higher",
"tolerance": 0.01,
"cluster_by": "conversation",
"description": "guardrail: strict recall does not drop by more than one point"
},
{
"id": "ndcg-10",
"metric": "ndcg_at_10",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "conversation"
},
{
"id": "latency",
"metric": "latency_ms",
"gate": "exploratory",
"direction": "lower",
"cluster_by": "conversation"
}
],
"estimate_usd": 6
},
{
"id": "locomo-dev",
"kind": "memory-qa",
"benchmark": "locomo",
"split": "dev",
"embed": "real",
"limit": null,
"search_pins": {
"search.mode": "balanced",
"search.reranker.enabled": "false",
"search.autocut": "false"
},
"top_k": 10,
"id_field": "id",
"exclude_when": [
"abstention=true"
],
"min_clusters": 3,
"comparisons": [
{
"id": "recall-all-5",
"metric": "recall_all_at_5",
"gate": "noninferiority",
"direction": "higher",
"tolerance": 0.01,
"cluster_by": "conversation",
"description": "guardrail: strict recall does not drop by more than one point"
},
{
"id": "ndcg-10",
"metric": "ndcg_at_10",
"gate": "exploratory",
"direction": "higher",
"cluster_by": "conversation"
},
{
"id": "latency",
"metric": "latency_ms",
"gate": "exploratory",
"direction": "lower",
"cluster_by": "conversation"
}
],
"estimate_usd": 2
}
],
"alpha": 0.05,
"seed": 42,
"draws": 10000,
"budget_usd": 8,
"created_at": "2026-10-04T18:45:19.479Z",
"notes": "Use-attributed feedback weights: the sequential replay-with-feedback mode lands in milestone M2; until then these are guardrails only. LME-S trimmed to a 100-question category-stratified subset (selectQuestions, seed 42) to stay inside the plan E3 cap; embeddings warm in the content-addressed cache so both arms read identical vectors."
}