RES — PUMA empirical results (canonical index)
Headline reconciled figures (reference hardware = NVIDIA RTX 2060, seed=42, temp=0.0):
| quantity | canonical value | note |
|---|---|---|
| Triage F1-macro | 0.5867 (N=200) | qwen2.5:3b, contextual-anchoring; zero-shot = 0.3898 |
| Best estimation MAE | 1.86 SP @ qwen2.5:7b | qwen2.5:3b = 2.91; zero-shot baseline = 5.7150 |
| H2 baseline MAE | 2.03 SP | predictor constant = 4.04 SP (mean SP, not a MAE) |
| H1 Wilcoxon | p ≈ 6.9×10⁻¹³, r = −0.51 | reject H0₁ |
| H2 Wilcoxon | p = 0.104, r = −0.16 | fail to reject H0₂ |
| Executed sweep | 27 configs / 2,700 inferences | 9 models × 3 scenarios × 1 strategy, N=100 |
| Energy / carbon | ≈68 Wh / ≈11.75 gCO₂eq | full 27-run sweep |
Source: PUMA project documentation + reconciliation verdict · Confidence: verified-at-primary-source