RES — A cold-vs-warm drift of up to 0.006 F1-macro points was observed between…

A cold-vs-warm drift of up to 0.006 F1-macro points was observed between the first (cold) run and later warm runs — far below experimental significance — hypothesized to relate to quantized-weight load order and FP effects, recommending always measuring in warm state after a warm-up run.

Source: PUMA project documentation · Traceability: corpus unit MEM-075 · Confidence: verified-at-primary-source