PN — Evaluated model: gemma4:e2b
gemma4:e2b — family Gemma 4 (Google), GGUF Q4_K_M, run locally via Ollama. Excluded from valid results: parse_failure_rate ≥0.98 across scenarios; it consumed ~41 Wh / 7.11 gCO₂eq (≈60.5% of the sweep’s emissions) while producing no usable output — the project’s canonical sunk-cost finding.
Source: PUMA project documentation + raw benchmark results · Traceability: ANXAG-015,PUMA-023 · Confidence: verified-at-primary-source