bgz-tensor: D-PALETTE-NNUE CORRECTED — FT columns ARE a palette256 tenant (GREEN, certified codec)#681
Conversation
…nant (GREEN) Corrects the 2026-07-11 scalar-256 FENCE, which used the wrong codec (hand-rolled k-means, not the certified Fisher-z cosine-replacement) and the wrong method (materialized: reconstruct weights + re-run eval, when a similarity/ranking claim is read off the distance without reconstruction). encoding-ecosystem.md — the mandatory codec map — was skipped. Corrected probe reuses the REAL codec, bgz_tensor::fisher_z::FisherZTable (Fisher-z i8, per-family 3σ gamma; certified ρ≥0.999). On 256 FT columns from nn-1b6a82263149 (32640 off-diagonal pairs, the Jina-v5 cert setup), pairwise cosine ranking read off the i8 table, no materialization: ρ_all = 0.99971 (Pearson 0.99993), cosine MAE 0.00655, 64 KB one-table-read ρ_mid = 0.99873 on the hard near-orthogonal band (|cos| ≤ 0.3, 19126 pairs) Both clear the ρ≥0.999 anchor. Fisher-z (arctanh) is monotone → rank-preserving by construction; only the i8 3σ quantization is lossy, negligibly. Verdict: the FT columns ARE a palette256 tenant (one-table-read similarity preserved). Scope: FT-column pairwise-cosine SIMILARITY, not byte-exact eval from codes. examples/nnue_palette_cosine.rs reads the FT-column fixture exported by stockfish-rs export_ft_columns. Board: doc D-SF-V3-2 flipped to GREEN; EPIPHANIES E-PALETTE-NNUE-COSINE-GREEN-1 (corrects/supersedes the scalar fence). Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Awg6TXocHcwTtc6eGsHcdD
|
Warning Review limit reached
Next review available in: 19 minutes Enable usage-based reviews in Billing to review now. Otherwise, wait until the next included review is available. How can I continue?After more reviews become available, a review can be triggered using the To avoid repeated limits, reduce automatic review volume by pausing incremental auto-reviews earlier, using label-based review opt-in, excluding WIP or generated PR titles, or requesting reviews manually when the PR is ready. If your team needs uninterrupted high-volume reviews, an organization admin can enable usage-based reviews. How do review limits work?CodeRabbit enforces per-developer PR review limits for each organization. Most developers receive the normal plan review availability. For paid Pro and Pro+ PR reviews, CodeRabbit uses adaptive limits for sustained high-volume activity. When a developer's recent PR review activity reaches the 95th percentile or higher among CodeRabbit users, additional reviews become available more gradually as earlier reviews age out of the rolling window. Please refer docs for additional details. Review details⚙️ Run configurationConfiguration used: Organization UI Review profile: CHILL Plan: Pro Plus Run ID: 📒 Files selected for processing (3)
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. Comment |
Bugbot couldn't run - usage limit reachedBugbot is counted against Cursor usage for this user or team, and this run hit a usage or spend limit. A user or team admin can review and increase usage limits in the Cursor dashboard. (requestId: serverGenReqId_c7e5efc6-e12d-45c3-9879-7e044a1c0ce6) |
![chatgpt-codex-connector[bot]](https://avatars.githubusercontent.com/in/1144995?s=60&v=4){kind=small}
There was a problem hiding this comment.
💡 Codex Review
Here are some automated review suggestions for this pull request.
Reviewed commit: ae58658253
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".
If Codex has suggestions, it will comment; otherwise it will react with 👍.
Codex can also answer questions or update the PR. Try commenting "@codex address that feedback".
Two review fixes on the D-PALETTE-NNUE probe gate: 1. A missing hard cut (rho_mid NaN, < 8 near-orthogonal pairs) no longer reports GREEN — it withholds GREEN and returns non-zero (INSUFFICIENT), so a degenerate export can't certify the tenant without the fine-discrimination bar exercised. 2. A fenced/below-threshold result now returns a non-zero exit code instead of exiting 0, so automation can distinguish a certified tenant from a failed gate. The exit code is now the gate: 0 only on GREEN; non-zero on fenced/insufficient. The input-absent path stays exit 0 (CI-safe). No number changed — real fixture still GREEN (rho_all 0.99971, rho_mid 0.99873), exit 0. Co-Authored-By: Claude <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01Awg6TXocHcwTtc6eGsHcdD
2 participants
Summary
Corrects the palette-tenant probe. The earlier scalar-256 "FENCED" result used the wrong codec (a hand-rolled k-means codebook, not the workspace's certified Fisher-z cosine-replacement) and the wrong method (materialized — reconstructed weights and re-ran eval, when a similarity/ranking claim is read off the distance without reconstruction).
encoding-ecosystem.md— the mandatory codec map — was skipped. Re-run with the real codec flips the verdict.Corrected result — GREEN
Reuses
bgz_tensor::fisher_z::FisherZTable(Fisher-z i8, per-family 3σ gamma; certified ρ≥0.999 on 21 Qwen3-TTS roles / 256 Jina-v5 centroids). On 256 FT columns fromnn-1b6a82263149(32,640 off-diagonal pairs — the Jina-v5 cert setup), pairwise cosine ranking read off the i8 table, no materialization:Both clear the certified ρ≥0.999 anchor. Fisher-z (
arctanh) is monotone → rank-preserving by construction; only the i8 3σ quantization is lossy, and it is negligible. Verdict: the FT columns ARE a palette256 tenant — one-table-read cosine similarity preserved.Scope (honest)
Proven for FT-column pairwise-cosine similarity (the one-table-read tenant claim). It does not claim byte-exact eval from palette codes — the palette is a similarity/routing tenant, not an exact-eval substitute (that remains the raw net's job). The two are non-contradictory: similarity ranking is preserved; absolute-eval reconstruction was never the palette's job.
Changes
crates/bgz-tensor/examples/nnue_palette_cosine.rs— the corrected probe (reads the FT-column fixture exported bystockfish-rs examples/export_ft_columns.rs; the codec drags arrow via holograph, so a file is the clean cross-crate seam)..claude/knowledge/stockfish-nnue-as-perturbation-cascade.md— D-SF-V3-2 flipped to GREEN with the certified-codec result + a correction note explaining why the first cut was wrong..claude/board/EPIPHANIES.md—E-PALETTE-NNUE-COSINE-GREEN-1prepended (corrects/supersedes the scalar-fence entry; standing lesson banked: consultencoding-ecosystem.mdand reuse the certified codec; measure ranking in distance space, never reconstruct).Testing
cargo run --manifest-path crates/bgz-tensor/Cargo.toml --release --example nnue_palette_cosine→ GREEN (shown above). fmt + clippy clean (-D warnings).🤖 Generated with Claude Code
https://claude.ai/code/session_01Awg6TXocHcwTtc6eGsHcdD
Generated by Claude Code