-
-
Notifications
You must be signed in to change notification settings - Fork 0
ceiling aware feedback adaptation gate
新規 feedback-adaptation experiment で、baseline relation MRR が理論上限の 1.0 に達して strict improvement を示せない場合、採否をどのように判定するか。
新規 experiment の primary relation gate は baseline relation MRR により分岐する。
- baseline relation MRR が 1.0 未満なら、treatment は strict improvement を示さなければならない。
- baseline relation MRR が 1.0 なら、treatment は non-regression を示さなければならない。
- どちらの場合も、endpoint/type projected path、direct lexical / directional-negative controls、credited-only mutation、deterministic replay、contamination、immutable output を含む全 safety gate を通過しなければならない。
ceiling case の non-regression pass は、追加の順位改善を示せないが劣化せず全 safety gate を満たしたことを記録する。これは generalization、default変更、production採用を意味しない。
- depends on none
- refines none
- supersedes none
- conflicts with none
- informs none
#31 の独立 holdout では control と treatment の relation MRR がともに 1.0 で、全 safety gate を通過した。一方、従来の strict-improvement-only gate では ceiling のため追加改善が不可能であった。この判断は、将来の新規 experiment の採否規則を ceiling-aware にするものであり、#31 の観測済み結論を遡及的に pass または adopted へ変更しない。
- この規則は結果観測前に固定する新規 experiment にだけ適用する。baseline relation MRR が 1.0 未満のときに strict improvement を non-regression へ緩和しない。
- #31 と #29 の fixture、gold、manifest、gate、result、実験文書を変更、再実行、再集計しない。
- existing default、API、router、production behavior を変更しない。
- ceiling pass は frozen minimal corpus と固定protocolの範囲に限り、一般 corpus やproduction qualityへの主張に拡張しない。
新規 feedback-adaptation experiment は baseline relation MRR が 1.0 未満なら strict improvement、1.0 なら全 safety gate を伴う non-regression を採用条件とする。ceiling case を pass と記録しても、default変更、generalization、production採用の根拠にはしない。
- Issue #35 - ceiling-aware gate の承認済み policy scope。
- Issue #31 - independent reproduction と ceiling observation の根拠。
- Issue #29 - prior feedback-adaptation experiment の不変境界。
- Requirements - 新規 experiment に適用する requirement。
- Decision Structure index - entry format、ownership、lifecycle。
- Decision Structure
- Observation lifecycle test exception
- Ceiling-aware feedback adaptation gate
- Single-corpus real feedback validation
- Repository-native controlled corpus
- Longitudinal feedback adaptation
- github-rag-mcp replacement compatibility
- Evidence-gated local feedback reinforcement
- Confirmed outcome feedback reinforcement
- Real-task shadow feedback validation
- Frozen output round-trip integrity