The Three Levers of QAT: A Mechanism Taxonomy for Quantization-Aware Robustness, and What a Passive Lattice Does Not Buy

Quantization-aware training is usually reported as an outcome. This paper takes a released QAT checkpoint apart against its regular sibling and asks what it actually buys. On one model pair (Gemma-4-12B instruction-tuned, regular vs the qat-q4_0-unquantized release) under lattice simulation only — block-32 symmetric round-to-nearest, no serialized artifact — the protection decomposes into three separately measurable levers. Lattice fit is real and format-specific (median relative int4 error 0.7131x on the trained block-32 format, 0.9481x per-channel, 0.9291x at int3 — with the ABSOLUTE error-norm ratios 0.3803/0.4992, the dose gap that drives two corrections below). Substrate flatness is real but modest (matched-absolute-dose error-vector transplants give 0.7785 and 0.8501). Error-direction quietness is a modest measured factor at both depths, not a dominant carrier: sealed norm-matched cells read 0.793 (int3, registered partial band 0.70-0.90; symmetric cell 0.880; int4 analogue 0.859); an unmatched int3 reading of 0.53, once presented as dominant, was predominantly the dose gap and is retired (Correction 7). The usual explanation is refuted in the scope the instrument supports: AVERAGE NEXT-TOKEN margins are ~23% SHALLOWER unquantized (0.8919 vs 1.1556 nats), with matched-fact armor untested; retention at identical perturbation is 2.3x better, a count statistic reported descriptively beside the mean-storm figure (0.7785) without a mechanism claim. Registered corrections print at full weight: a "substrate-carried at 0.48" reading retracted as dose-confounded after an external referee catch, and the direction demotion above — leaving a product of modest factors with no dominant carrier. Edit-shaped generalization is RETIRED with its sign inverted (Correction 8): a per-checkpoint-normed reading (median 0.6984) was dose-confounded (the QAT arm received 0.457x the absolute dose); at shared absolute dose, matched to =6/8 bar — a one-seed pre-registered bar-miss, not a demonstrated equivalence. What survives is the bar-miss and an exact set-identity replication across the paired screen, consistent with — but not proving — the working hypothesis that the mechanism lives in the training loop. Scope is stated plainly: one checkpoint pair for the taxonomy, one model and one seed for the null, an open-set probe, closed-set panel counts as lower bounds (this program measured the open-set channel at 1.79x the closed-set figure), and no claim meeting this program's >=3-family / >=2-size bar. The comparison is release against release, so the vendor-recipe confound is open throughout: the two checkpoints differ by the vendor's entire QAT pipeline rather than by a controlled intervention, and every "QAT bought X" should be read as "this QAT release differs from its sibling in X." Small-model and single-pair phenomena generating frontier predictions, not findings about deployed systems. Research conducted in close collaboration with an AI assistant (Claude, Anthropic) under a sealed pre-registration discipline; disclosure and contribution statement in the paper.