navi-SAD¶
A dynamical systems probe for LLM inference.
Runs softmax and linear attention in parallel on the same frozen weights, measures per-head cosine divergence, and reconstructs the model's internal attractor via delay-coordinate embedding.
Gate Status¶
| Gate | What | Status | Last verified |
|---|---|---|---|
| 0 | Non-interference --- identical tokens + logits with/without hooks | Pass | 2026-03-24 |
| 1 | Parity --- recomputed fp32 softmax through o_proj matches native | Pass | 2026-03-24 |
| 2 | Stability --- 50 generations, zero VRAM creep | Pass | 2026-03-24 |
| 3 | Rank correlation --- per-head permutation entropy (PE) vs known fractal dimension | In progress | --- |
What's New¶
2026-04-27 --- TruthfulQA case closed. 400-sample length-matched permutation null returned p=0.96; the dense-d direction was a generation-length confound, not signal. Instrument framing pivots to dynamical-systems probe with no application claims pre-Gate-3.
2026-03-25 --- Pilot findings published. 40-sample TruthfulQA pilot: grand-mean SAD does not separate groups; per-head PE recurrence count surfaced (later killed by 400-sample replication).
2026-03-24 --- Gate 2 passes. 50 consecutive generations, zero VRAM creep.
2026-03-24 --- Gates 0 and 1 pass. Instrument verified on Mistral-7B.
Documentation¶
| Section | Contents |
|---|---|
| Getting Started | Install, run the gates |
| Theory | What SAD measures and why --- Takens' embedding, capacity gap, adjacent literature |
| Instrument | Gate discipline, validation proofs, adapter rules |
| Research | Pilot findings, roadmap, open problems |
| How-To | Contributing guide |
| Reference | Module reference, frozen decisions, glossary, changelog |