Quantization damage is multiplicative, not additive — and your eval suite averages over exactly the decisions that break. @holistic_ai + @ai_ucl (Zekun Wu, Swati Dhiman, Adriano Koshiyama) measured the decision itself instead of the benchmark score. The damage is silent: at 3 bits an agent stops calling its tools. One bit lower, it loses roughly half its safety refusals. Benchmarks barely move. The mechanism: quantization does not add fixed-size noise to a decision margin. Quantization multiplies the margin by a survival factor. The factor collapses with bit-width: median 0.86 at 4 bits, 0.33 at 3, 0.00 at 2. That shrinkage removes the protection a large margin gave. The model's own small bias then picks the failure direction. In the example model, 96% of should-call decisions flip to inaction at 3 bits, while which-tool selection stays intact. The evidence: 16 models, 8 families, quantized with round-to-nearest, GPTQ, and llama.cpp GGUF builds, 8 down to 2 bits. The fitted law predicts flip rates within a median of 1.8 percentage points on held-out decisions. Per-decision flip probabilities are calibrated: 0.004 expected calibration error over 131,758 predictions. The production playbook: the law's constants are per-model and do not transfer. Measure a small paired margin set for your model at your bit-width. Under the paper's cost-matched tests, nothing repairs damage more cheaply than one more bit. arxiv.org/abs/2608.06564