docs: calibration session handover — H1-H5 hypotheses, Cronbach α, full protocol 5 testable hypotheses: H1: BF16 truncation flips ~5% of ranks (bucket boundary effect) H2: γ+φ encoding preserves more rank order than linear CDF H3: i8 signed > u8 unsigned for gate-heavy roles specifically H4: ICC profile correction brings ALL encoding paths to ρ > 0.998 H5: Cronbach α reveals which tasks need multi-lens vs single lens Testing protocol: Phase 1: ONNX f32 ground truth (rten, Jina v5, 1000 pairs) Phase 2: BF16 baseline (stream GGUF, same CLAM) Phase 3: 5 encoding paths (linear, γ+φ, i8, γ+φ signed, spiral) Phase 4: Spearman ρ before/after ICC per path Phase 5: Cronbach α across 6 lenses Synthesis matrix: which encoding × which role × ICC or not. Estimated: 3-4 hours. Validates everything built in 67+ commits. https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp#114
Merged
Conversation
BF16 7-bit mantissa flips rank order for cosines within ±0.008. Spearman ρ drops ~5% from BF16 truncation alone, not encoding. Calibrate against ONNX f32 to isolate pure encoding error. ICC profile then corrects BOTH encoding AND BF16 truncation. Camera analogy: calibrate lens against RAW, never against JPEG. https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp
When raw cosine is within ±0.008 of a HEEL bucket boundary, BF16 truncation can flip the bucket assignment. High precision refinement (HIP/TWIG) on the wrong bucket = confidently lost. Fix: boundary_risk metadata per centroid pair. 95% safe → fast cascade 5% uncertain → skip cascade, validate at LEAF or compute directly γ+φ golden ratio stride reduces boundary risk by placing bucket edges at irrational positions that don't align with BF16 quant steps. https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp
…ll protocol 5 testable hypotheses: H1: BF16 truncation flips ~5% of ranks (bucket boundary effect) H2: γ+φ encoding preserves more rank order than linear CDF H3: i8 signed > u8 unsigned for gate-heavy roles specifically H4: ICC profile correction brings ALL encoding paths to ρ > 0.998 H5: Cronbach α reveals which tasks need multi-lens vs single lens Testing protocol: Phase 1: ONNX f32 ground truth (rten, Jina v5, 1000 pairs) Phase 2: BF16 baseline (stream GGUF, same CLAM) Phase 3: 5 encoding paths (linear, γ+φ, i8, γ+φ signed, spiral) Phase 4: Spearman ρ before/after ICC per path Phase 5: Cronbach α across 6 lenses Synthesis matrix: which encoding × which role × ICC or not. Estimated: 3-4 hours. Validates everything built in 67+ commits. https://claude.ai/code/session_01ChLvBfpJS8dQhHxRD4pYNp
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.