PN — Model scaling is non-monotonic on these PMO tasks: in 4 of 6 (family x…
Model scaling is non-monotonic on these PMO tasks: in 4 of 6 (family x scenario) cells larger models did not improve results (e.g. gemma3:1b beats gemma3:12b on prioritization 0.87 vs 0.57; qwen2.5:3b beats 7b on triage), supporting that small local LLMs are competitive on narrow domain tasks.
Source: puma docs/results/phase_b_analysis.md · Traceability: corpus unit
PUMA-024· Confidence: repo-backed