Hypothesis — expB_token_stability#

Consumes expA's block-energy trace. Decides the temporal-locality route for C1's escalation bytes (and G01/G06/G12/G13 cache designs): is the important- block set stable enough across consecutive tokens to cache?

Hypothesis
  The per-token important-block set (blocks covering 95% of FFN energy) is
  temporally sticky: mean Jaccard(t, t+1) ≥ 0.5, decaying slowly with
  distance, and the UNION working set over a 128-token generation stays
  well below the full model (≤80% of blocks) — i.e., a generation has a
  reusable working set that a RAM cache can hold, so residual/expert bytes
  are fetched once per window, not once per token.

Falsification criterion
  If Jaccard(t, t+1) < 0.3 (set churns almost completely every token) or
  the 128-token union ≥ 95% of all blocks (no working set exists at
  generation scale), the temporal-locality route (b) for C1 dies, leaving
  only batch amortization — C1 then merges into C2 (amortized
  verification), and G06/G12 prefetchers must predict, not cache.

Method
  Load expA's block_energy_trace.npz (positions × 28 layers × 96 blocks,
  with domain and trajectory ids). Per trajectory and layer:
  top-set(t) = smallest block set covering 95% of position-t energy.
  Metrics: mean Jaccard(t, t+δ) for δ ∈ {1, 2, 4, 8, 16, 32}; union
  working-set fraction over windows {8, 32, 128} tokens; per-layer and
  per-domain aggregates. Cross-prompt preview (expC-lite): per-trajectory
  aggregate top-set, Jaccard within-domain vs across-domain.

Baseline
  Random sets of identical size (expected Jaccard for independent draws)
  — the null; and δ→∞ behavior (unconditional overlap between distant
  tokens) as the floor stickiness must beat.

Result
  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).

Interpretation
  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).

Next experiment
  Prototype margin-gated deferred-refinement runtime; expG rerun at 8B.

Analysis — expB_token_stability#

Run: results/expB_token_stability/20260812T045224Z/ · consumes expA's trace (Qwen3-1.7B, 6 144 positions × 28 layers), 64-neuron sets covering 95% of per-token FFN energy.

Hypothesis / Falsification
  Hoped: Jaccard(t,t+1) ≥ 0.5 with slow decay; 128-token union ≤80%.
  Kill: Jaccard(t,t+1) < 0.3 OR 128-token union ≥95%.

Result — KILL CRITERION TRIGGERED (union), stickiness ILLUSORY
  Jaccard(t,t+δ): 0.690 (δ=1) → 0.688 (δ=16) → 0.686 (δ=32) — flat.
  Random-set null at matched sizes: 0.640. Real stickiness beyond
  set-size artifact: +0.05 — negligible. (The registered ≥0.5 threshold
  is technically met at 0.69, but the null exposes it as a size effect:
  sets covering ~77% of blocks overlap ~64% by chance alone.)
  Union working set: 98.0% of blocks after 8 tokens, 99.3% after 32,
  99.9% after 128 — the entire FFN is touched within a few tokens.
  Domain locality (expC-lite): within-domain = across-domain = 1.0 —
  every trajectory's union covers everything; no domain-specific caches.

Interpretation
  1. NEGATIVE, decisively: for dense SwiGLU models, per-token FFN energy
     working sets are neither sparse at pageable granularity (expA) nor
     temporally reusable (here) nor domain-clustered (here). The
     DejaVu / LLM-in-a-flash paradigm does NOT transfer to modern
     non-ReLU architectures without ReLUfication retraining — now
     measured, not just suspected from the literature. G06 (TealPager),
     G12 (HiddenPrefetch for dense FFNs) and G13 (DomainCache) are dead
     on this model family; G07 (WandaTiers) survives only for STATIC
     (input-independent) importance tiering.
  2. Scope note (honesty): we measured activation-ENERGY sets on a 1.7B
     dense model. Residual-relevance sets (which bytes fix a decision)
     could differ, and MoE models have architectural, not emergent,
     sparsity — C3 is untouched by this result. But the prior for any
     per-token dynamic weight selection on dense SwiGLU is now strongly
     unfavorable.
  3. PIVOT (with expF, expA): all three fine-grained routes for cutting
     C1's per-token escalation bytes are closed — layers (expF), blocks
     (expA), temporal caching (expB). What remains is BATCH AMORTIZATION,
     which expH makes attractive: the full 1 GB residual streams
     sequentially in ~80 ms at 13 GB/s; a margin-gated queue that defers
     low-margin tokens and refines them in periodic sweeps shares one
     sequential residual pass across many tokens. C1 thereby converges
     with C2 (Amortized Verification Sweeps) — the evidence has selected
     the candidate.

Next experiment
  Prototype candidate: margin-gated deferred-refinement runtime
  (C1→C2 merge): 4-bit resident base generates optimistically; low-margin
  tokens queue; a periodic sequential residual sweep verifies/corrects
  (speculative-decoding-style rollback on flips). Measure end-to-end
  bytes/token, tok/s, and agreement on this Mac. Also rerun expG at 8B to
  check margin-signal scaling before committing the prototype design.

expB_token_stability#

Stability of important weight blocks across consecutive tokens (Jaccard, transitions, working-set lifetime)

Status: scaffolded 2026-08-11, not yet run.