RES — Gender-prefix robustness testing on triage found qwen2.5:3b shifts…

Gender-prefix robustness testing on triage found qwen2.5:3b shifts accuracy only 3-4pp under a gendered name prefix while qwen2.5:1.5b shifts 11-12pp (~25% of predictions flip); effects below ~5pp are not reliably distinguishable from noise at N=100.

Source: puma docs/results/bias_evaluation.md · Traceability: corpus unit PUMA-043 · Confidence: repo-backed