spb/localvm-research Public License

1---2project: localvm-research3document: expB_token_stability/hypothesis4author: Simon-Pierre Boucher5contact: contact@spboucher.ai6created: 2026-08-127modified: 2026-08-128status: reviewed9---1011# Hypothesis — expB_token_stability1213Consumes expA's block-energy trace. Decides the temporal-locality route for14C1's escalation bytes (and G01/G06/G12/G13 cache designs): is the important-15block set stable enough across consecutive tokens to cache?1617```text18Hypothesis19  The per-token important-block set (blocks covering 95% of FFN energy) is20  temporally sticky: mean Jaccard(t, t+1) ≥ 0.5, decaying slowly with21  distance, and the UNION working set over a 128-token generation stays22  well below the full model (≤80% of blocks) — i.e., a generation has a23  reusable working set that a RAM cache can hold, so residual/expert bytes24  are fetched once per window, not once per token.2526Falsification criterion27  If Jaccard(t, t+1) < 0.3 (set churns almost completely every token) or28  the 128-token union ≥ 95% of all blocks (no working set exists at29  generation scale), the temporal-locality route (b) for C1 dies, leaving30  only batch amortization — C1 then merges into C2 (amortized31  verification), and G06/G12 prefetchers must predict, not cache.3233Method34  Load expA's block_energy_trace.npz (positions × 28 layers × 96 blocks,35  with domain and trajectory ids). Per trajectory and layer:36  top-set(t) = smallest block set covering 95% of position-t energy.37  Metrics: mean Jaccard(t, t+δ) for δ ∈ {1, 2, 4, 8, 16, 32}; union38  working-set fraction over windows {8, 32, 128} tokens; per-layer and39  per-domain aggregates. Cross-prompt preview (expC-lite): per-trajectory40  aggregate top-set, Jaccard within-domain vs across-domain.4142Baseline43  Random sets of identical size (expected Jaccard for independent draws)44  — the null; and δ→∞ behavior (unconditional overlap between distant45  tokens) as the floor stickiness must beat.4647Result48  KILL TRIGGERED: 128-token union = 99.9% of blocks (≥95% line); Jaccard(t,t+1)=0.69 vs random null 0.64 — stickiness is a set-size artifact. Domain locality nil (within=across=1.0).4950Interpretation51  No cacheable working set for dense SwiGLU energy sets: DejaVu/LLM-in-a-flash paradigm does not transfer without ReLUfication. G06/G12/G13 dead on this family; pivot to batch amortization (C1 merges into C2).5253Next experiment54  Prototype margin-gated deferred-refinement runtime; expG rerun at 8B.55```56