quantcompose

mixedPythonInference & servingJuly 11, 2026

the question

Do per-layer quantization errors compose additively, as GPTQ/AWQ assume?

what came out

The additive-sensitivity premise holds from Q8_0 to Q4_K_M (A within CI of 1) but leans superadditive at 2-bit (A=1.13): isolated per-layer sensitivity under-counts the damage at aggressive bit-widths. Exact full-vocab KL oracle; reverses the pre-registration.

method & receipts

Result: mixed
Pre-registered — the prediction was written down and committed before the run.

→ read the code and re-run it

github.com/v-code01/quantcompose