spb/localvm-research Public License

1---2project: localvm-research3document: expF_error_accumulation/analysis4author: Simon-Pierre Boucher5contact: contact@spboucher.ai6created: 2026-08-127status: reviewed8---910# Analysis — expF_error_accumulation1112Run: `results/expF_error_accumulation/20260812T043945Z/` · code committed before13run. Qwen3-1.7B, 28 layers in 7 depth-groups of 4, affine g64 4-bit, 4814trajectories × 128 tokens teacher-forced, 17 configurations.1516```text17Hypothesis / Falsification18  Hypothesized ≥3× sensitivity spread across depth groups and ≥40% of lost19  agreement recovered by repairing the top ~25% of layers.20  Kill criteria: <2× spread, or top-25% repair recovering <20%.2122Result — HYPOTHESIS REFUTED (spread criterion killed; repair borderline-negative)23  All-4-bit floor: 87.53% agreement (loss = 12.47 points vs bf16).24  DEGRADE-ONE (one group 4-bit, rest bf16): agreement drops span only25    3.7–5.6 points across the 7 groups — a 1.5× spread (< the 2× kill26    line). Mid-depth groups (L12–19) are marginally most sensitive; the27    LAST group is among the LEAST sensitive (3.96) despite expD showing28    the largest hidden-state error there — late-layer error is large in29    norm but decision-benign.30  REPAIR-ONE (one group bf16, rest 4-bit): best single group (L12–15)31    recovers +1.5 points = 12% of the loss; worst 5%.32  REPAIR-TOP-K: top-2 groups (29% of layers) recover 24% of the loss;33    top-3 (43% of layers) recover 33% — consistently SUB-proportional.34  Cross-check of additivity: individual degrade-one drops sum to ~3335    points, yet degrading everything at once costs only 12.5 — errors36    partially mask each other; symmetrically, repair values sum to 7.437    of the 12.5 lost — repair requires cooperation across depth.3839Interpretation40  1. NEGATIVE RESULT (recorded per charter §10/§17): quantization damage41     to token decisions is DIFFUSE and cooperative across depth, not42     concentrated. Layer-restricted escalation cannot materially cut43     bytes-per-escalation: paying 29% of the residual bytes buys only44     24% of the repair — worse than linear, no leverage.45  2. expD's depth-concentrated hidden-state error was a red herring for46     decision repair: large late-layer representation drift coexists with47     benign decisions (norms ≠ decisions — a caution for any design that48     gates on hidden-state error instead of decision margin).49  3. Consequently C1's bytes-per-escalation must come from the remaining50     mechanisms: (b) temporal locality — the residual working set of51     escalated tokens staying hot in RAM (expB); (c) sub-layer/block-level52     selection — repair only the weight blocks that matter for THIS token53     (expA/expE, finer grain than layers); or (d) batch-amortized54     escalation — G17/G18-style verification sweeps sharing one residual55     read across many queued low-margin tokens.56  4. Design implication: if (b) also fails, C1 degrades into "resident57     4-bit + rare whole-model refinement passes" — which is exactly58     candidate C2 (amortized verification). The two candidates are59     converging on the same mechanism from opposite ends; this is60     useful, not disappointing.6162Next experiment63  expA (weight contribution concentration at BLOCK granularity within64  layers) — the finer-grained version of the question expF just answered65  negatively at layer granularity; feeds expE (partial GEMM) directly.66  In parallel, expB (temporal stability of important blocks) decides the67  page-cache path (b).68```69