Hypothesis — expF_error_accumulation#

Follows expD: escalation currently means touching the FULL residual; if layer sensitivity is concentrated, escalation can be restricted to a subset of layers and candidate C1's bytes-per-escalation drops proportionally.

Hypothesis
  Layer sensitivity to quantization error is strongly non-uniform: degrading
  a single depth-group to 4-bit (rest bf16) hurts agreement unevenly across
  groups (≥3× spread between most and least sensitive), and symmetrically,
  repairing only the most sensitive ~25% of layers (bf16 in-group, 4-bit
  elsewhere) recovers a disproportionate share — ≥40% — of the agreement
  lost by the all-4-bit model.

Falsification criterion
  If per-group degradation effects are near-uniform (<2× spread), or if
  repairing the best 25% of layers recovers <20% of the lost agreement
  (i.e., error is diffuse and cooperative across depth), then
  layer-restricted escalation cannot cut bytes-per-escalation materially
  and C1 must rely entirely on temporal locality (expB) or block-level
  selection (expE).

Method
  Qwen3-1.7B bf16 reference, same 48-trajectory teacher-forced protocol
  (benchmarks/datasets/eval_prompts.json, 128 tokens, greedy reference).
  28 transformer layers → 7 contiguous depth groups of 4.
  (i) DEGRADE-ONE: quantize (affine g64, 4-bit) all divisible Linear
      layers of one group; rest bf16. 7 runs → sensitivity map.
  (ii) REPAIR-ONE: all layers 4-bit except one group at bf16. 7 runs →
      repair-value map, plus the all-4-bit floor (from expD stage0 B).
  (iii) REPAIR-TOP-K: bf16 for the k most-repairing groups (k=1,2), 4-bit
      elsewhere → cumulative repair curve vs bytes.
  Metrics per config: agreement with reference, mean KL. Embedding/head
  layers excluded (kept bf16 throughout, as in expD/expG).

Baseline
  All-bf16 (agreement=1 by construction) and all-4-bit (87.5%, expD
  stage0 B) bracket every configuration. No straw men.

Result
  REFUTED. Sensitivity spread across depth groups: 1.5× (kill line: <2×).
  Top-2 groups (29% of layers) repair only 24% of lost agreement —
  sub-proportional; repair values sum to 7.4 of 12.5 points lost (repair
  is cooperative, not concentrated). Full numbers:
  results/expF_error_accumulation/20260812T043945Z/ and analysis.md.

Interpretation
  Quantization damage to decisions is diffuse across depth. Layer-
  restricted escalation gives no leverage; expD's late-layer hidden-state
  error is decision-benign (norms ≠ decisions). C1's bytes-per-escalation
  must come from temporal locality (expB), block-level selection
  (expA/expE), or batch amortization (→ convergence with C2).

Next experiment
  expA (block-granularity concentration) + expB (temporal stability).

Analysis — expF_error_accumulation#

Run: results/expF_error_accumulation/20260812T043945Z/ · code committed before run. Qwen3-1.7B, 28 layers in 7 depth-groups of 4, affine g64 4-bit, 48 trajectories × 128 tokens teacher-forced, 17 configurations.

Hypothesis / Falsification
  Hypothesized ≥3× sensitivity spread across depth groups and ≥40% of lost
  agreement recovered by repairing the top ~25% of layers.
  Kill criteria: <2× spread, or top-25% repair recovering <20%.

Result — HYPOTHESIS REFUTED (spread criterion killed; repair borderline-negative)
  All-4-bit floor: 87.53% agreement (loss = 12.47 points vs bf16).
  DEGRADE-ONE (one group 4-bit, rest bf16): agreement drops span only
    3.7–5.6 points across the 7 groups — a 1.5× spread (< the 2× kill
    line). Mid-depth groups (L12–19) are marginally most sensitive; the
    LAST group is among the LEAST sensitive (3.96) despite expD showing
    the largest hidden-state error there — late-layer error is large in
    norm but decision-benign.
  REPAIR-ONE (one group bf16, rest 4-bit): best single group (L12–15)
    recovers +1.5 points = 12% of the loss; worst 5%.
  REPAIR-TOP-K: top-2 groups (29% of layers) recover 24% of the loss;
    top-3 (43% of layers) recover 33% — consistently SUB-proportional.
  Cross-check of additivity: individual degrade-one drops sum to ~33
    points, yet degrading everything at once costs only 12.5 — errors
    partially mask each other; symmetrically, repair values sum to 7.4
    of the 12.5 lost — repair requires cooperation across depth.

Interpretation
  1. NEGATIVE RESULT (recorded per charter §10/§17): quantization damage
     to token decisions is DIFFUSE and cooperative across depth, not
     concentrated. Layer-restricted escalation cannot materially cut
     bytes-per-escalation: paying 29% of the residual bytes buys only
     24% of the repair — worse than linear, no leverage.
  2. expD's depth-concentrated hidden-state error was a red herring for
     decision repair: large late-layer representation drift coexists with
     benign decisions (norms ≠ decisions — a caution for any design that
     gates on hidden-state error instead of decision margin).
  3. Consequently C1's bytes-per-escalation must come from the remaining
     mechanisms: (b) temporal locality — the residual working set of
     escalated tokens staying hot in RAM (expB); (c) sub-layer/block-level
     selection — repair only the weight blocks that matter for THIS token
     (expA/expE, finer grain than layers); or (d) batch-amortized
     escalation — G17/G18-style verification sweeps sharing one residual
     read across many queued low-margin tokens.
  4. Design implication: if (b) also fails, C1 degrades into "resident
     4-bit + rare whole-model refinement passes" — which is exactly
     candidate C2 (amortized verification). The two candidates are
     converging on the same mechanism from opposite ends; this is
     useful, not disappointing.

Next experiment
  expA (weight contribution concentration at BLOCK granularity within
  layers) — the finer-grained version of the question expF just answered
  negatively at layer granularity; feeds expE (partial GEMM) directly.
  In parallel, expB (temporal stability of important blocks) decides the
  page-cache path (b).

expF_error_accumulation#

Error accumulation: which layers tolerate, amplify, or recover from controlled approximation

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