[AdaWorldAPI]

Purpose

Preserve already executed TANH proof receipts on the mainline board, without falsely marking the future native D-TFLP-1 integration as implemented. Record the new research boundary for adaptive/learning-time calibration, specifically the existing 8-byte FamilyGamma wire payload, and nominate Tract typed graph/cooking as the next cold architectural harvest, consistent with D-NAM and D-ONNX-INTAKE-0.

This is a docs-only evidence/index PR. No code, kernels, importers, models, receipts or learning loop added; no old research branch rewritten.

Files

• .claude/board/entries/2026-10-26-d-tflp-tanh-proof-learning-tract-harvest.md: durable index with detailed links to original experimental results and explicit follow-on test gates.
• .claude/board/STATUS_BOARD.md: one additive crosslink above the stage table, with D-TFLP-1 still Queued/NOT RUN for production-native integration.
• .claude/board/entries/2026-10-26-d-tflp-65-landing.md: correct probes(D-TFLP 6.5): softmax fixture replay + certified preparation gate #1800/probes(D-TFLP 6.5): Codec8 semantic admission + DT64/Quack/D-NAM handoff #1801 from drafts to merged and DT64-QB to canonical D-DT64-Q0; link TANH evidence.

Research anchors, not copied or reexecuted

Source branch claude/bpe-tokenization-architecture-3xd4eh; frozen research commit 4082dbb79; original parity audit Amendments 3/6.1/6.2/6.3, TANH/LOGISTIC source-vs-math results, fusion evidence.

• TANH (test scope): 2,224 calibration × 256 codes = 569,344 evaluated. SourceParity matches pinned TFLM 569,344/569,344; distinct interval-certified CertifiedMath matches an independent high-precision math reference 569,344/569,344. 17 source-vs-math differences were found and attributed to the fixed-point exp polynomial near quantization boundaries. This is NOT a universal source=math identity; rather, both clearly delimited artifacts have passed their own oracles.
• Composition: TANH → LOGISTIC prepared fusion matches two TFLM stages on 569,344 tested outputs when intermediate quantization is retained; intentionally dropping it produces 69,652 differences. Falsifiers are cited.
• Related: merged #1800 Softmax numerical evidence and #1801 Codec8 admission. FamilyGamma-z softmax algebraically removes TANH from that path; its parity does not independently certify a numerical TANH kernel.

The new, still OPEN work

1. Learning-time calibration and generations: changing weights/activations/quantizer statistics invalidates prepared-table receipts absent a new witnessed admission; source and math contracts must remain distinct. OBSERVE → PREHEAT → PROPOSE → VALIDATE → COMMIT is a proposal, not a trained system.
2. 8-byte FamilyGamma: bgz-tensor/src/fisher_z.rs stores only z_min and z_range as two little-endian f32 values. Existing persisted bytes have no version or generation field, and -128 is a valid code here, not SymNaN. Test re-calibration, real held-out model families, stale generation, persistence compatibility, saturation and training semantics before any format/version change. Do not mutate the 8 bytes.
3. Tract first harvest priority: upstream InferenceModel → TypedModel → declutter/optimize → execute separates typed operator/graph transformations from physical execution, making it a strong cold frontend and algebra/representation separation reference. #1802 retains the objective Burn onnx-ir comparison. No Tract runtime in native hot paths, no automatic importer adoption, no whole-model parity claims.

Owners and limits

D-NAM / EXO-152 numerical method/generation; D-NFC #1753; #1798 typed numerical leaf and #1799 canonical source-attested D-DT64-Q0; #1802 cold ONNX frontend bake-off.

Tests for this PR: NOT RUN, documentation changes only. Existing test results are cited historical evidence, not newly performed work. No merge without review.

[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: 348a128d-8b60-429a-bd99-f9858145d6a0

• 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 2026-10-26T22:13:45.687355Z	5b47047	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.