[AdaWorldAPI]

What

The pre-registered probe for EXO-129 (TORQUE T-2026-10-23-1). Pre-registration was committed before any probe code (124d8e9). The results were appended afterwards and the registered rules were not changed.

It asks whether a 256×256 table learned from observed one-hop transitions composes over k hops where ComposeTable does not (P-S2, #1726). It also asks whether a data-only purity statistic predicts how well a table composes.

Results

Agreement with the truth at k = 2 for the learned table (L) and for ComposeTable (C):

setting	purity	L k=2	C k=2
W1 clustered	0.471	0.017	0.029
W2 uniform	0.479	0.009	0.027
W3 XOR group, n = 0 / 500 / 1500	1.000	1.000	1.000
W3 XOR group, n = 3000	0.544	0.145	0.260

• Torque: KILL. L is below the 0.90 bar in both W1 and W2, and it does worse than C there.
• Mechanism M: PASS by the registered rule (Spearman 0.935), but the evidence is thin. Four of the six settings tie, so the rank rests on three distinct points. The magnitude prediction (agreement ≈ purity^k) fails.
• Harness control and both disables hold:
  • the shuffled table reads ≤ 0.005;
  • setting L := C reproduces C exactly.
• Both verdicts are pinned. Disable run: inverting purity fails the M assertion.

Reading, not a ruling: purity below 1 means the cell does not determine the next cell, so no one-byte table can compose. In Markov-chain terms that is a lumpability failure, and it is a candidate name for crumb trail T-CLOSURE (EXO-132). W3 with n = 3000 also shows that a closed carrier composes with its exact law even when snapping is noisy: C stays flat at about 0.25 across k = 1..6, while the learned table decays.

Files

• crates/bgz-tensor/tests/exo129_learned_lut_probe.rs, with three tests: the probe with pinned verdicts, the D2 disable, and a check that W3 is an XOR group.
• .claude/board/entries/2026-10-23-exo129-learned-lut-preregistration.md, holding the pre-registration plus the results, plus the regenerated entries index.

Checks

cargo test --release --test exo129_learned_lut_probe passes all 3 tests in 2.7 s, built with debug info off. cargo clippy --release --tests reports no warnings for the probe.

The ledger updates (TORQUE row KILLED, crumb 5 and the naming attempt) go on #1729's branch, where those files live.

---

Generated by Claude Code

[coderabbitai]

Important

• 🔍 Trigger review

This repository does not receive automatic reviews because it has fewer than 10 stars.

⚙️ Run configuration

• Source mode: Auto
• Repository: Unavailable
• Branch: Unavailable
• File: No configuration file from this repository applied
• Commit: Unavailable
• Layers: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: 56967e4a-790e-466f-b184-f89bec5d26bf

• 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.

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{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 2026-10-23T08:49:21.029605Z	b56bb56	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.