Repository navigation
probe(palette): falsify multi-hop error containment with closed and open codebooks - #1721
Conversation
|
Important
This repository does not receive automatic reviews because it has fewer than 10 stars. ⚙️ Run configuration
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. Comment |
Moves crates/bgz-tensor/tests/palette_multihop_error_probe.rs here from #1721. Unlike the local helpers already in this series, it exercises the real implementation, so the series keeps its experiment contract and gains the concrete counter-probe. #1721 is closed as superseded once this lands. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
Codex Review SummaryThis comment shows the latest Codex review activity on this pull request.
ℹ️ About Codex in GitHubYour team has set up Codex to review pull requests in this repo. Reviews are triggered when you
Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings. |
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01G22yT6htkcdyXsihxxXdrg
Purpose
Falsify (rather than assume) the proposal that an immutable Palette256 and a pairwise 256² relation/compose LUT structurally prevent multi-hop error explosion and hallucinations. Freeze the reference frame and compare actual
ComposeTablelookups against exact Base17 XOR composition for 1..64 hops.What's already shipped
bgz_tensor::ComposeTableandBase17::xor_bindprovide a real pairwise composition and an independent scalar oracle. The immutable reference-frame decision and Fisher-ZFamilyGammarelation codec are separate; Spearman ρ≈0.9998 is pairwise rank fidelity on a tested sample, not a hop guarantee.New executable falsifiers
Next measured study
Real held-out codebooks and typed relations, hop count 1/2/4/8/16/32/64, exact-domain oracle, codebook closure, relative/absolute error, top-k disagreement, bootstrap uncertainty, source overlap, failure-to-abstain, null controls. Compare direct immutable-address lookup vs explicit re-quantization and exact f32. Multi-hop drift/hallucination prevention remains UNPROVEN.
Verification
cargo test --manifest-path crates/bgz-tensor/Cargo.toml --test palette_multihop_error_probe.No hot-path, ABI, composition, or codebook changes. No Google Search implementation ported.