Refuse recursive tracker proposals without paired baseline wins

.agents/pm/epics/pm-rl-apvf.toon

@@ -6,10 +6,10 @@ status: open
 priority: 1
 tags[3]: execution,recursive-training,roadmap
 created_at: "2026-09-07T21:21:56.236Z"
-updated_at: "2026-10-24T21:47:38.159Z"
+updated_at: "2026-10-25T02:10:05.514Z"
 author: codex-pm-ecosystem
 acceptance_criteria: Two real training generations produce changed checkpoints and collect with the promoted successor; a bounded LLM adapter demonstrates parameter updates; two controllers launch a job once; crash and cancellation drills preserve artifacts and budget integrity; contamination and invalid checkpoints stop promotion; every authored source passes exact four-metric coverage and documented-surface gates
-comments[40]{created_at,author,text}:
+comments[41]{created_at,author,text}:
   "2026-09-07T23:06:06.595Z",codex-pm-native,"Execution foundation is implemented under child pm-rl-s44k: a real bounded TypeScript contextual-bandit policy-gradient adapter supplies promoted checkpoints to successor collection. RECURSIVE_TRAINING.md distinguishes this numerical adapter from the remaining SDK-persisted job controller, atomic approved-budget reservations, owner leases, crash/cancellation recovery, Compose isolation and LLM parameter-update acceptance. The epic remains open. Repeated evaluation is adaptive validation, so a separate final benchmark is still required."
   "2026-09-10T22:02:44.990Z",pi-agent,"Promotion gate design (closing the recursive self-improvement loop):\n\nPROPERTY: A candidate policy is promoted ONLY when its held-out expected reward exceeds the incumbent's by at least the effect threshold, with confidence >= 1-alpha, using at least minSamples evaluation episodes, AND the held-out evaluation set is not reachable from the candidate's training data. The default answer is NO: absent, unreadable, self-referential, incomparable, or statistically insufficient evidence promotes NOTHING.\n\nDOMAIN: Pairs of bounded-reward evaluation evidence records (candidate + incumbent) carrying a finite empirical mean, a finite sample count, and declared reward bounds [lo, hi]; a confidence level in (0,1), a minimum sample count >= 1, an effect threshold >= 0; and a contamination verdict (null = isolated).\n\nSTATISTICAL CRITERION: Hoeffding inequality for bounded random variables. Distribution-free, valid for any bounded objective a caller pins a range to, conservative for small samples (the property a fail-closed gate wants). epsilon = (hi - lo) * sqrt(ln(2/alpha)/(2n)). Promote iff candidate lower confidence bound exceeds incumbent upper confidence bound by the effect threshold (maximize; inverted for minimize). Chosen over a distribution-assuming bound because RL reward distributions are unknown and a single lucky sample must not promote; Hoeffding needs no distribution assumption and is loosest exactly when samples are few.\n\nWHAT IT BUYS: with probability >= 1-alpha, a promoted candidate's TRUE expected reward exceeds the incumbent's by at least the effect threshold. A single lucky sample, a regression, a contaminated held-out set, and a sub-threshold improvement are all refused.\n\nWHAT IT DOES NOT BUY: protection against a held-out set semantically duplicated under a different name (the contamination gate's job, consumed not rebuilt); non-stationary environments; unbiased estimates after repeated selection (adaptive validation, not a final benchmark); reward hacking on the held-out set itself.\n\nHELD-OUT ISOLATION IS STRUCTURAL: the bandit adapter validates training and evaluation example identities disjoint before any gradient, so the held-out set is unreachable from the training loop by construction; the gate's contamination verdict is null. The persisted promoteGeneration flow already enforces the same via findContaminationPath over provenance edges.\n\nLINEAGE: every promotion and every refusal is recorded. In the recursive loop (bandit adapter) each generation receipt carries promoted + refusalReason + the sampled held-out means; the generations array is the auditable chain from generation 0 to N. Transitive invalidation uses the existing buildLineageAncestry forward propagation and invalidate.ts (not rebuilt).\n\nREMAINING ON THIS EPIC: the durable execution controller (Docker Compose trainer, atomic budget reservations, owner leases, crash/cancellation recovery, LLM parameter-update acceptance) and wiring the statistical gate into the persisted rl generation promote command's scores contract remain separate slices. This delivery closes the recursive self-improvement loop's promotion decision; it does not deliver the full durable controller."
   "2026-09-12T23:19:05.412Z",pi-agent,"Audit 2026-09-13: EXISTS — the bounded TypeScript contextual-bandit adapter is implemented and merged (bandit.ts, PR #33, pm-rl-s44k now closed); the fail-closed statistically bounded promotion gate is implemented and merged (promotion.ts, PR #35, pm-rl-gr7j closed); contamination and invalid checkpoints stop promotion (promotion.test.ts); 323 tests pass with 100/100/100/100 coverage across 20 source files. REMAINS — the durable execution controller is not implemented: no Docker Compose trainer adapter, no atomic budget reservations, no owner leases, no crash/cancellation recovery drills; a bounded LLM adapter demonstrating parameter updates is not implemented (the bandit is a numerical adapter, not an LLM trainer); two controllers launching a job once is not demonstrated. These are the remaining slices named in the item description and the 2026-09-10 pi-agent comment."
@@ -50,7 +50,8 @@ comments[40]{created_at,author,text}:
   "2026-10-22T10:31:49.788Z",codex-agent,"Complete CI verified at 31b9f7b996c9192c9230: https://github.com/unbraind/pm-rl/actions/runs/39761523871 . Independently downloaded all three completed job logs: npm on Node 22, npm on Node 26, and Bun on Node 26 each pass 690/690, zero fail/skip/cancel/todo, 100/100/100/100 coverage over 48 sources. Strict tracker health, unscoped lint, zero duplication, 340 documented declarations, attestation, production audit, packing, changelog, and Bun graph installation pass; CodeQL passes. The local duplicate full gates were interrupted under host load, not counted as passes. This receipt update changes tracker evidence only; executable source and tests remain identical. Claim is released; leave the epic in progress. No new pm CLI/SDK defect found."
   "2026-10-22T10:32:17.818Z",codex-agent,"Receipt correction: the complete verified CI head is 31b9f7b5520eef7b1a55cb7f1863d9692ad81b38. The extra characters after the valid 31b9f7b prefix in the preceding comment were a transcription error. CI run 39761523871 and all recorded counts are verified against this exact head."
   "2026-10-24T21:47:38.159Z",codex-agent,"Returned to open in the 2026-10-24 sweep: Open PR https://github.com/unbraind/pm-rl/pull/95 carries remaining recursive-promotion work; multiple narrower merged implementations do not establish all execution, contamination, crash and resource criteria in one verified delivery. No active agent is established."
-files[16]:
+  "2026-10-25T02:09:56.051Z",codex-agent,"PR 95 promotion-gate infrastructure remains tracked by child pm-rl-baseline. Owner-authorized main-based history restoration replaces this branch divergent epic projection with the exact main item/history, then re-records branch intent through the pinned CLI. The bounded title-vector candidate must beat the incumbent and both constants with item-paired bootstrap lower effects above 0.01 plus all original gates. Fresh public corpus: 105 training, 41 held-out, constant 0.8780487805, candidate rewards 0.8713980699 / 0.8645003749 / 0.8534392510, zero promotions. Resume preserves all checkpoints and charges. These are infrastructure and conditional validation results, not a claimed self-improvement win."
+files[18]:
   - path: bandit.ts
     scope: project
   - path: corpus.ts
@@ -59,6 +60,8 @@ files[16]:
     scope: project
   - path: index.ts
     scope: project
+  - path: item-evaluation.ts
+    scope: project
   - path: loop.ts
     scope: project
   - path: package.json
@@ -73,6 +76,8 @@ files[16]:
     scope: project
   - path: test/bandit.test.ts
     scope: project
+  - path: test/baseline.test.ts
+    scope: project
   - path: test/corpus.test.ts
     scope: project
     note: Snapshot regression and complete publication snapshots
@@ -86,7 +91,7 @@ files[16]:
     scope: project
   - path: test/promotion.test.ts
     scope: project
-tests[7]:
+tests[8]:
   - command: npm test
     path: test/promotion.test.ts
     scope: project
@@ -153,7 +158,21 @@ tests[7]:
       created_at: "2026-10-22T09:35:47.428Z"
       source_kind: local_mutation
       source_ref: feat/recursive-training-generations
-docs[4]:
+  - command: "node --test --test-name-pattern=\"paired gate property|vector capacity|constant\" test/baseline.test.ts"
+    scope: project
+    timeout_seconds: 600
+    pm_context_mode: none
+    workspace_context_mode: snapshot
+    provenance:
+      author: codex-agent
+      created_at: "2026-10-25T02:10:02.481Z"
+      source_kind: local_mutation
+      source_ref: feat/rsi-beats-baseline
+docs[6]:
+  - path: docs/baseline-frontier-evidence.json
+    scope: project
+  - path: docs/baseline-frontier.md
+    scope: project
   - path: docs/process-trainer.md
     scope: project
   - path: docs/tracker-loop-evidence.json