spb/localvm-research Public License

1---2project: localvm-research3document: expF_error_accumulation/hypothesis4author: Simon-Pierre Boucher5contact: contact@spboucher.ai6created: 2026-08-127modified: 2026-08-128status: reviewed9---1011# Hypothesis — expF_error_accumulation1213Follows expD: escalation currently means touching the FULL residual; if layer14sensitivity is concentrated, escalation can be restricted to a subset of15layers and candidate C1's bytes-per-escalation drops proportionally.1617```text18Hypothesis19  Layer sensitivity to quantization error is strongly non-uniform: degrading20  a single depth-group to 4-bit (rest bf16) hurts agreement unevenly across21  groups (≥3× spread between most and least sensitive), and symmetrically,22  repairing only the most sensitive ~25% of layers (bf16 in-group, 4-bit23  elsewhere) recovers a disproportionate share — ≥40% — of the agreement24  lost by the all-4-bit model.2526Falsification criterion27  If per-group degradation effects are near-uniform (<2× spread), or if28  repairing the best 25% of layers recovers <20% of the lost agreement29  (i.e., error is diffuse and cooperative across depth), then30  layer-restricted escalation cannot cut bytes-per-escalation materially31  and C1 must rely entirely on temporal locality (expB) or block-level32  selection (expE).3334Method35  Qwen3-1.7B bf16 reference, same 48-trajectory teacher-forced protocol36  (benchmarks/datasets/eval_prompts.json, 128 tokens, greedy reference).37  28 transformer layers → 7 contiguous depth groups of 4.38  (i) DEGRADE-ONE: quantize (affine g64, 4-bit) all divisible Linear39      layers of one group; rest bf16. 7 runs → sensitivity map.40  (ii) REPAIR-ONE: all layers 4-bit except one group at bf16. 7 runs →41      repair-value map, plus the all-4-bit floor (from expD stage0 B).42  (iii) REPAIR-TOP-K: bf16 for the k most-repairing groups (k=1,2), 4-bit43      elsewhere → cumulative repair curve vs bytes.44  Metrics per config: agreement with reference, mean KL. Embedding/head45  layers excluded (kept bf16 throughout, as in expD/expG).4647Baseline48  All-bf16 (agreement=1 by construction) and all-4-bit (87.5%, expD49  stage0 B) bracket every configuration. No straw men.5051Result52  REFUTED. Sensitivity spread across depth groups: 1.5× (kill line: <2×).53  Top-2 groups (29% of layers) repair only 24% of lost agreement —54  sub-proportional; repair values sum to 7.4 of 12.5 points lost (repair55  is cooperative, not concentrated). Full numbers:56  results/expF_error_accumulation/20260812T043945Z/ and analysis.md.5758Interpretation59  Quantization damage to decisions is diffuse across depth. Layer-60  restricted escalation gives no leverage; expD's late-layer hidden-state61  error is decision-benign (norms ≠ decisions). C1's bytes-per-escalation62  must come from temporal locality (expB), block-level selection63  (expA/expE), or batch amortization (→ convergence with C2).6465Next experiment66  expA (block-granularity concentration) + expB (temporal stability).67```68