gemma4-quant-regime-study / CLAIMS_MATRIX.md
moebiusT7's picture
v1.0.1 public package + figures + paper PDF (DOI 10.5281/zenodo.21875513)
21bbd75 verified
|
Raw
History Blame Contribute Delete
4.39 kB

Claims matrix

Labels: Supported = supported within the frozen local scope only; Partially supported = directionally relevant but incomplete; Unsupported = data reject or do not establish; Requires replication = needs a new condition. All inferential claims are exploratory (no preregistration, single run per cell).

Candidate claim Status Exact evidence / boundary
No cell separates by an operationally large margin on Core-500 Supported as description Spread 0.022 (11 tasks of 500). NOT a tie claim: no equivalence test was run, and 12B-PTQ vs 26B-PTQ is nominally significant (17 vs 6 discordant, p=0.035, uncorrected)
The cells are statistically equivalent ("tied") Unsupported No TOST/equivalence test anywhere; one cross-model McNemar rejects at 0.05
QAT improves the clean dense 12B pair Unsupported Paired Δ −0.012 (QAT-only 1 vs PTQ-only 7), p=0.0703; Smoke agrees in sign (−0.020)
QAT harms the clean dense 12B pair Partially supported Same evidence; marginal, single-run, clustered rows — directional only
QAT improves the released 26B MoE pair Partially supported Paired Δ +0.018 (15 vs 6), p=0.0784; Smoke agrees (+0.030); pair impure (base revision / expert config unverifiable / offload confound)
The regime × released-pair interaction is nonzero Supported as exploratory finding Delta-of-deltas +0.030; stem-clustered bootstrap 95% CI [+0.010, +0.052] (primary; 100 stems × 5 paraphrases) and task-iid CI [+0.008, +0.052]; single-run, no multiplicity correction
The interaction is a pure QAT-recipe (or architecture) effect Unsupported 26B pair confounds regime with possible base revision, unverifiable expert config, and a CPU-offload compute path; only the 12B pair is same-base, all-GPU
The interaction concentrates in volatile_current Supported as description Crossed family pattern 0.5314/0.4971 vs 0.4971/0.5314 (93↔87 of 175); the numeric identity of the crossed counts is coincidence; no test attached
Quantization regime moves safety-critical failure rates Unsupported 0.000 in 7 of 8 runs; one over-verification event in 26B-PTQ Core-500 (0.008 = 1/125). Rare-event floor: rule-of-three 95% upper bounds ≈0.6–2.4 %/run
num_ctx pinning (QAT profiles) is behavioral Unsupported (controlled) 12B-PTQ rerun at num_ctx 8192: 100/100 byte-identical vs default ctx. Control run on 12B only; extrapolated to 26B
Regime choice preserves exact temperature-0 outputs Unsupported Cross-regime byte-identity 23.2–32.0 %; median first divergence 46–73 chars
Observed cross-regime divergence is runtime nondeterminism Unsupported (Smoke-100) Same-configuration repeats 100/100 in ALL FOUR cells, incl. the CPU-offloaded 26B-PTQ, across run windows. Core-500 attribution is inference (no Core-500 repeats); a systematic offload contribution to 26B rows cannot be excluded
QAT 26B remains a defensible production binding Supported as operational inference Best-or-tied on volatile_current (0.5314) and stale_premise (0.9600), fastest memory-fitting cell (1,613 ms); partially circular w.r.t. the same-day binding decision (two cells reused) — the anti-diagonal is the independent evidence
PTQ 12b should remain the dense-tier rollback binding Supported as operational inference Clean-pair QAT effect ≤ 0 (p=0.07, directional); no reason to change the rollback target
date_boundary degrades under QAT Partially supported −0.02 in both pairs (n=50 each); descriptive only
Latency differs by regime Unsupported (by design) 26B-PTQ column is CPU-offload-dominated (26/74 split); no cross-memory-boundary latency claim is made
Results generalize to other runtimes / stacks / content-quality benchmarks Requires replication One GPU, Ollama 0.30.6, one governance stack, routing tasks only
The models have uncontaminated scores Unsupported Public weights; no training-corpus audit; suite prompts local and unpublished at run time
The study is reproducible on the same artifacts Supported Pinned digests and hashes; 13 runs / 2,900 rows / 0 errored; every routing statistic incl. the family table recomputes from row level via the deposited script (families from the suites' authoritative family fields); safety rates quoted from archived harness scorer outputs