eval: add longitudinal feedback adaptation experiment

README.md

@@ -243,6 +243,8 @@ Trace-credited feedback adaptationは、relation traceの`record_success`が後
 
 独立reproductionはprior feedback-adaptation resultを選択入力にせず、新規D1 development / holdout splitで同じtrace-credit claimを検証します。relation pathはruntime fieldを除き、`source_id`、`target_id`、`edge_type`だけへ射影してgoldと比較します。prior fixtureは識別子だけのcontamination auditに使い、prior goldとresultは読みません。詳細は[Trace-credited feedback adaptation reproduction experiment](docs/feedback-adaptation-reproduction-experiment.md)を参照してください。
 
+Longitudinal multi-corpus feedback adaptation fixes `h=0,1,2,3` across three disjoint clusters in each split. Control records the same feedback schedule without edge mutation, while treatment reinforces only credited relation paths. Headroom clusters require strict final-horizon aggregate MRR improvement; ceiling clusters require non-regression only. Earlier observed feedback-adaptation results are never selection inputs. See [Longitudinal multi-corpus feedback adaptation experiment](docs/longitudinal-feedback-experiment.md).
+
 ## Explanation model
 
 各 `SearchHit` は次の情報を保持します。

docs/longitudinal-feedback-experiment.md

@@ -0,0 +1,44 @@
+# Longitudinal multi-corpus feedback adaptation experiment
+
+## Purpose
+
+This frozen experiment tests whether trace-credited feedback remains directionally useful after multiple success events across independently constructed corpus clusters. It compares a no-mutation control with trace-credited treatment at fixed horizons `h=0,1,2,3`.
+
+It does not reuse, rerun, modify, aggregate, or use observed results from the earlier feedback-adaptation experiments as selection inputs. Their fixture identifiers are read only by the contamination audit; their gold and observed result artifacts are not loaded.
+
+## Frozen cohorts and provenance
+
+- Manifest: `tests/fixtures/d1_liplus_longitudinal_feedback_experiment.manifest.json`
+- Each development and holdout split contains three disjoint clusters: two headroom clusters and one ceiling cluster.
+- Every cluster has a separately hashed fixture, gold schedule, and provenance record. The records identify these as deterministic in-repository constructed corpora, with no network or D1 access and zero writes.
+- The audit compares all six clusters and both earlier feedback-adaptation fixture families by document path, node ID, source URL, normalized query, endpoint, and complete edge identity. It reads prior fixture identifiers only.
+
+## Schedule and path identity
+
+Every arm shares the same corpus, configuration, limit, timestamped feedback schedule, and scoring schedule. At horizons one through three, both arms record the same successful relation feedback event. The control records it without an edge mutation; treatment uses only the credited relation trace path.
+
+Relation path checks retain raw observed steps for audit but compare only the projected identity:
+
+```json
+{"source_id":"...","target_id":"...","edge_type":"..."}
+```
+
+Runtime fields cannot affect identity. Each horizon also checks a direct lexical case and a directional-negative case.
+
+## Ceiling-aware gate and one-shot execution
+
+For headroom clusters, the control relation MRR at `h=0` must be below `1.0`; treatment must strictly improve final-horizon aggregate MRR and must not regress in any headroom cluster. For ceiling clusters, the `h=0` control relation MRR must equal `1.0`, and treatment requires non-regression only. A ceiling pass records that no additional rank improvement is evidenced; it never authorizes default changes, generalization, or production adoption.
+
+After the result-free freeze is pushed, development is allowed once at its registered output path:
+
+```powershell
+python tools/run_longitudinal_feedback_experiment.py development `
+  --manifest tests/fixtures/d1_liplus_longitudinal_feedback_experiment.manifest.json `
+  --output tests/fixtures/d1_liplus_longitudinal_feedback_experiment.development.result.json
+```
+
+Only a passing development result allows one holdout execution at its registered output path. The runner rejects any other output path and refuses to overwrite an observed result. Once observation begins, fixture, gold, provenance, audit, manifest, implementation, tests, documentation, result aggregation, and reruns are forbidden.
+
+## Interpretation boundary
+
+Even a passing frozen result supports only longitudinal relation retrieval adaptation on these six small constructed clusters. It does not evaluate agent task success, tool calls, tokens, latency, autonomous-feedback correctness, a production default, or general corpora.

docs/requirements.md

@@ -94,6 +94,7 @@
 76. feedback adaptationのdevelopment / holdout、feedback event、score query、expected path、gate、registered run count、hash、contamination audit、exclusive outputを結果観測前に固定し、development全gate通過時だけholdoutを一度開く。
 77. feedback adaptation reproductionはprior resultを選択入力にせず、prior fixtureとの識別子だけのcontamination auditを行う。raw relation stepは比較前に`source_id`、`target_id`、`edge_type`だけへ射影し、runtime fieldを含むsynthetic testでpath identityをfreeze前に検証する。
 78. 新規 feedback-adaptation experiment の primary relation gate は、baseline relation MRR が 1.0 未満ならtreatmentのstrict improvement、baseline relation MRR が 1.0 ならtreatmentのnon-regressionを要求する。いずれも endpoint/type projected path、direct lexical / directional-negative controls、credited-only mutation、deterministic replay、contamination、immutable output を含む全 safety gate を必須とする。ceiling case のnon-regression pass は追加の順位改善を示せないことを記録するだけで、generalization、default変更、production採用を許可しない。
+79. longitudinal feedback adaptation experiment は development / holdout の各3以上の相互disjoint corpus cluster で `h=0,1,2,3` を固定し、同一 feedback schedule の no-mutation control と credited-relation-path-only treatment を比較する。headroom は final-horizon aggregate MRR の strict improvement と cluster non-regression、ceiling は non-regression を要求し、endpoint/type-only path、controls、deterministic replay、contamination、exclusive immutable output を全horizonで必須とする。過去 feedback-adaptation result は selection input にせず、prior fixture identifier のみを contamination audit に使用する。
 75. v3 implementation、prompt、manifest、query override、schema、集約、path audit、hash規則、gate、stop rule、testsをresult-free commitでpushした後、development stage / 4 case packet / 12 responses / resultを各一度だけ生成する。
 76. development全12 gate通過時だけholdout stageを一度生成し、異なるfresh 12 judgesで同じgateを評価する。packet、response、resultの上書き、観測後の規則変更、実LLM品質値のCI再生成を拒否する。
 
@@ -125,3 +126,4 @@
 - [Node-first blind selection experiment](node-first-blind-selection-experiment.md) がsingle-case invocation、node-majority、channel provenance分離、v1 / v2不変監査、conditional holdoutを定義する。
 - [Trace-credited feedback adaptation experiment](feedback-adaptation-experiment.md) がcontrol / treatmentの因果比較、result-free freeze、conditional holdoutを定義する。
 - [Trace-credited feedback adaptation reproduction experiment](feedback-adaptation-reproduction-experiment.md) が新規D1 split、prior-result非参照、endpoint/type-only path projection、conditional holdoutを定義する。
+- [Longitudinal multi-corpus feedback adaptation experiment](longitudinal-feedback-experiment.md) が固定horizon、複数cluster、ceiling-aware gate、one-shot conditional holdoutを定義する。