KW — Reproducibility (warm-state)
Reproducibility (warm-state). Bit-exact identical task metrics across repeated runs after a warm-up, achieved via fixed seed=42, temperature=0.0, response caching and pinned versions.
Source: PUMA project documentation + raw benchmark results · Traceability: ANXAG-007,ANXAG-008,ANXAG-009 · Confidence: verified-at-primary-source