D-SELF-CALIBRATING-LAB-0 phases I–IV: G2 law-A boundary, P5 Beta correspondence, P1 certified argmax, lab v0 (Gomoku)

[AdaWorldAPI]

Phases I–IV of D-SELF-CALIBRATING-LAB-0. All changes are tests, probes and docs; no production default changes.

Phase I: P8 re-check, G2 boundary, board drift

• P8 re-checked in code. No production code builds a CutContext or calls replay_chain. replay_step writes revision's F/C over the truth that forward computed, so the emitted edge names the weight's opcode but carries a revision truth. Recorded, not changed.
• G2 decision (scoped). New decision-bearing wiring uses exact revision (law A). Existing defaults, including NarsEngine::new's build(1), stay as they are. G1 stays separate.
• New fixture a_load_bearing_fixture_under_law_a: law A goes 210 → 64 and flips the verdict. Only B16 agrees; B1–B8 do not. Disable (law A → B1): red.
• DEFAULT_FREQUENCY_BAR's rationale no longer cites B1's constant confidence 170 as evidence.
• STATUS_BOARD drift fixed (D-RPF-0, D-MOORE-NARS-0) and four missing rows added.

Phase II-a: P5, NARS ↔ Beta (causal-edge/tests/d_rpf_p5_beta.rs)

• E1: e is the Jeffreys (k = 1) posterior mean, exactly.
• E2: unit observations stall at confidence code 244 after 20 observations. Past the stall, CE64 revision behaves like a fixed-window estimate, not a conjugate Beta update; mean |Δe| exceeds 0.05 at n = 256.
• E2b: the stall code depends on observation strength (12-row table).
• E3: duplicated evidence counts as new. Four copies claim a posterior sd 2.04× too small.
• E4: a variance rebuilt from the codes is coarse at high confidence.
• E5: revision never reaches c = 255.
• Disable (ceil rounding): E2, E2b and E5 red.

Phase II-b: P1, certified argmax on Gomoku (minimax_certified)

• Deterministic certificate: a candidate stops once its running minimum is at or below the best finished value (L_best ≥ U_j).
• On 400 positions it picks the same move as minimax at every k, with 36 / 45 / 61 % fewer folds at k = 4/4, 5/5, 8/8.
• Disable (unsound cut at L + 1): changes a pick.

Phase IV: lab v0

examples/shared/lab.rs: paired comparison over matched blocks (exact sign-flip p, t-interval), Holm correction, pre-registration verdicts, reliability with ECE and Brier, and a certificate kind on every result. The Holm stop and the min-effect gate are each disable-verified. The first min-effect test was vacuous and has been fixed.

Gomoku arm (moore_nars_gomoku_probe --lab). A pilot exposed three design errors:

• H4 must be measured with the cost term off, because a cost-aware chooser reacts to the cheaper ASC;
• 100 games only measured the learning phase;
• calibration was read over the wrong population.

These were fixed before a confirmatory run on 16 fresh seed blocks:

hypothesis	result
H2 learned chooser beats the best fixed recipe	SUPPORTED (+0.024)
H4 certified minimax cuts folds	SUPPORTED (−29 folds/step; games identical at cost 0)
H7 fold-cost term cuts folds	SUPPORTED (−143 folds/step, at −0.026 score)
H3 delayed credit beats immediate	INCONCLUSIVE (survives alone, not Holm)
H5 decay adapts faster after a switch	FALSIFIED

Calibration: the chooser's expectation has ECE 0.05 over non-neutral outcomes and 0.32 over all outcomes.

Verification

• lance-graph-planner: 553 passed.
• causal-edge P5: 6 passed.
• cognitive-shader-driver, run as CI runs it: 437 passed, 0 failed.
• fmt clean. Pre-existing clippy findings in untouched files are left alone.

Not done here

• The crossword arm, the held-out Moore domain, and a SoA-shaped experiment record.
• Thompson, eligibility-trace and context-conditioned arms.
• Several stale "In PR" rows outside these areas (D-CE64-, D-LXC-, D-MBX-CUTOVER-0).

🤖 Generated with Claude Code

https://claude.ai/code/session_01EFw2WdKr1oxvaKCJC2ua2R

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Essentials
• Run ID: 79e4aaba-e31a-4cc9-bb03-0d3e2b409beb

---

📥 Commits

Reviewing files that changed from the base of the PR and between 6e40f43 and b448d88.

---

📒 Files selected for processing (12)

• .claude/board/STATUS_BOARD.md
• .claude/board/entries/2026-10-10-d-lab-p1-certified-argmax-gomoku.md
• .claude/board/entries/2026-10-10-d-rpf-g2-law-a-boundary.md
• .claude/board/entries/2026-10-10-d-rpf-p5-nars-beta-correspondence.md
• .claude/board/entries/2026-10-10-d-self-calibrating-lab-0.md
• .claude/board/entries/README.md
• crates/causal-edge/tests/d_rpf_p5_beta.rs
• crates/cognitive-shader-driver/examples/crossword_crossing_care_probe.rs
• crates/cognitive-shader-driver/examples/moore_nars_gomoku_probe.rs
• crates/cognitive-shader-driver/examples/shared/lab.rs
• crates/lance-graph-planner/src/chain_counterfactual.rs
• crates/lance-graph-planner/tests/d_rpf_table_1.rs

---

 _________________________________________________________________________________________
< The three laws of code review: 1. Find bugs. 2. Find more bugs. 3. Find even more bugs. >
 -----------------------------------------------------------------------------------------
  \
   \   (\__/)
       (•ㅅ•)
       / 　 づ

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

---

• Autofix · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

Warning

Billing warning: we have not been able to collect payment for this subscription for more than 72 hours. Please update the payment method or pay any pending invoices in Billing to avoid service interruption.

---

_{Comment @coderabbitai help to get the list of available commands.}

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-10T09:43:46.845842Z	b448d88	Draft marked ready

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.