validate: f32 engine on real tables — attractor collapse persists F32 tables don't fix collapse. Root cause: ReLU kills inhibition. Qwen3-VL: 40% diversity, entropy +1.0% (diffuses) Jina-v5: 20% diversity, entropy +3.3% (diffuses) Reranker: 40% diversity, entropy +2.1% (diffuses) Even with f32 precision, the MatVec cycle converges to dominant eigenvector because ReLU(negative) = 0 removes all inhibition. Next: remove ReLU, use softmax normalization instead, or use signed accumulation without clamping negatives to zero. https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A#141
Merged
Conversation
FINDINGS (BF16 tables, 256 centroids, 20 queries × 3 models): - Thinking INCREASES entropy (diffuses, doesn't focus) - Top-5 overlap with plain cosine: 6-24% (nearly uncorrelated) - Kendall tau: 0.01-0.14 (no rank agreement) - 60x slower than one-shot cosine lookup ROOT CAUSE: MatVec with positive-only cosines = power iteration → converges to dominant eigenvector regardless of input → same attractor wins for every query = attractor collapse The iterative cycle needs real inhibition (signed tables with negative values) or it's just expensive diffusion. Lane calibration test also included: BF16: Spearman ρ=0.9999, Pearson r=0.9999 (PERFECT) i8 direct: ρ=0.997, r=0.999 (EXCELLENT) u8 CDF: ρ=0.999, r=0.80 (RANK ok, VALUES destroyed) γ+φ: ρ=1.000 vs CDF (IDENTICAL ordering, γ+φ is a no-op) https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A
Stage 0: Foundations (DONE) — empirical results, what works/doesn't Stage 1: Fix table (f32 primary, signed i8 from real cosines) Stage 2: Contrastive table learning (forward pass → table update) Stage 3: SPO 2³ causal certificates (8 octants × 28 layers) Stage 4: L1-L27 gate reward shaping (epiphany detection) Stage 5: L4 holographic memory (bundle fingerprint + replay) Stage 6: NARS reasoning (truth values + gap detection) Stage 7: OSINT pipeline (spider-rs + Reader-LM + embedding) Stage 8: 4096×16 branch graph (sparse, 250× faster than dense) Critical path: Stage 1→2→3→4 (10-15 days) 4 parallel tracks: Core Engine, Causal Reasoning, Memory, OSINT Risk register + metrics for each stage. https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A
…sts pass) Stage 1A: F32ThinkingEngine (f32_engine.rs) - Full-precision distance table, NO floor heuristic - Signed MatVec: accumulate → abs-normalize → ReLU → re-normalize - Temperature scaling: 1/T before normalization - from_codebook_f32(): compute pairwise cosine directly - 7 tests: roundtrip, convergence, diverse peaks, reset, top-k, temperature Stage 1B: Signed engine wiring fix (signed_engine.rs, builder.rs) - from_unsigned() marked DEPRECATED (CDF rank shift = fake sign) - builder.raw_cosines() → from_f32_cosines() (real cosine signs) - BuiltEngine::F32 variant in builder enum Stage 2A: ContrastiveLearner (contrastive_learner.rs) - Online table learning: table[i][j] += α × (real_cos - table[i][j]) - Fan-out: 1 forward pass → K neighbor updates - Symmetry maintained (both [i][j] and [j][i]) - LearnerStats tracking (update_count, MAE, alpha) - 4 tests: pair update, monotonic MAE decrease, symmetry, fan-out Also fixed pre-existing test compilation errors: - sensor.rs: CODEBOOK_SIZE import in test module - role_tables.rs: StackedN + f32_to_bf16 import in test helper 310 tests pass, 0 failures. https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A
F32 tables don't fix collapse. Root cause: ReLU kills inhibition. Qwen3-VL: 40% diversity, entropy +1.0% (diffuses) Jina-v5: 20% diversity, entropy +3.3% (diffuses) Reranker: 40% diversity, entropy +2.1% (diffuses) Even with f32 precision, the MatVec cycle converges to dominant eigenvector because ReLU(negative) = 0 removes all inhibition. Next: remove ReLU, use softmax normalization instead, or use signed accumulation without clamping negatives to zero. https://claude.ai/code/session_019RzHP8tpJu55ESTxhfUy1A
AdaWorldAPI
pushed a commit
that referenced
this pull request
May 13, 2026
…ed in ndarray master moved blake3 behind the hpc-extras feature gate as part of the i686/no_std unblock work (AdaWorldAPI/ndarray PR #141). lance-graph specifies default-features = false on its ndarray dep and previously only opted into 'std', which left blake3 unresolved at workspace build time even though src/hpc/{plane,seal,merkle_tree,vsa}.rs use it unconditionally. Adding 'hpc-extras' to the feature list restores the hash dep. lance-graph already uses every other HPC module from the fork (Fingerprint, CAM-PQ, CLAM, BLAS, ZeckF64), so opting into hpc-extras is the intended posture for this consumer. Fixes the linux-build, test, and test-with-coverage failures on PR #364.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
No description provided.