fix: Grammar/Markov hardening — slice unification, kernel wiring, parser tests, triangle distance#282
Conversation
…ation + normalize
…idates + cosine length assert
…ference param + per-slot has_unfillable
…orkers B-fix-slices changed slice() return type from (usize, usize) tuple to RoleKeySlice struct. Subsequent workers (B-fix-triangle, B-fix-parser via trajectory.rs tests, markov_bundle tests) used the old tuple destructure. This commit threads the new struct fields (slice.start, slice.stop) through all consumers. Tests: 289 pass on lance-graph-contract; 66 pass on deepnsm; 8 pass on the integration_role_alignment test (THE test that should have blocked PR #279).
![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: eeddfe97d2
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".
| let total_bits = (n * 64) as f32; | ||
| bits_diff as f32 / total_bits |
There was a problem hiding this comment.
hamming_normalized divides by 64 * n, but qualia_to_binary_fingerprint only sets 18 meaningful bits per word. With the current expected_qualia_footprint shape (vec![0u64; 1]), the maximum achievable classification distance is 18/64 = 0.28125, so the ticket router’s novel-domain gate (classification_distance > 0.7 in ticket_emit.rs) becomes unreachable when this distance is used, preventing Extrapolation routing in genuinely novel cases.
Useful? React with 👍 / 👎.
| for id in [ | ||
| // Pronouns + demonstratives (rank 0..30 in COCA) | ||
| 2, 4, 8, 12, 14, 18, 22, 26, 28, | ||
| // Common NSM-mapped function words (rank 30..200) | ||
| 35, 45, 58, 67, 73, 89, 102, 117, 134, 158, 192, | ||
| // Mental predicates | ||
| 201, 233, 287, 309, 354, |
There was a problem hiding this comment.
This table is currently an approximation with only a small subset of IDs, even though parser routing treats it as the authoritative NSM-prime detector. Because parse_with_coverage now relies entirely on is_nsm_prime, missing primes are counted as non-primes, which systematically depresses primes_found and can over-trigger Abduction (primes_found < 4) for otherwise semantically rich inputs.
Useful? React with 👍 / 👎.
2 participants
Summary
Addresses all 3 CRITICAL + 6 HIGH findings from PR #279 review:
markov_bundle.rsnow imports fromrole_keysexclusively (was equal-partition 16384/5 = 3277; canonical is domain-specific 2000/2000/2000/1500/1500). Integration testintegration_role_alignment.rs(THE test that should have blocked feat: DeepNSM grammar parser — Markov ±5 bundler, role keys, thinking styles #279) asserts alignment.rotate_rightremoved. Post-bundle rotation was corrupting role-slice alignment. Deleted entirely; documented for future per-sentence pre-bundle permute.coherence_at_with_kernel(i, kernel)andtotal_coherence_with_kernel(kernel)honour the kernel's weight function. Uniform vs MexicanHat produce different scores (tested).sum(|weights|)after accumulation; cosine comparisons now kernel-invariant.parse_with_coverage,coverage_failed,maybe_emit_ticket.r < 64heuristic with explicitNSM_PRIME_IDS: HashSet<u16>(COCA-mapped).compute_classification_distancenow does normalized Hamming over qualia fingerprint (was permanent 0.0 stub).role_candidatesparameterized. Explicitthreshold+top_kparameters; deprecated shim for old signature.attempted_inferenceparameterized. No longer hardcoded to Deduction; deprecated legacy shim.has_unfillable. Any-slot-None triggers CounterfactualSynthesis (was all-slots-None).Stats
Test plan
cargo check -p lance-graph-contract— greencargo check --manifest-path crates/deepnsm/Cargo.toml --all-features— greencargo test -p lance-graph-contract --lib— 289 passedcargo test --manifest-path crates/deepnsm/Cargo.toml --lib— 66 passedcargo test --manifest-path crates/deepnsm/Cargo.toml --test integration_role_alignment— 8 passedhttps://claude.ai/code/session_01SbYsmmbPf9YQuYbHZN52Zh