RES — PUMA empirical results (canonical index)

Headline reconciled figures (reference hardware = NVIDIA RTX 2060, seed=42, temp=0.0):

quantitycanonical valuenote
Triage F1-macro0.5867 (N=200)qwen2.5:3b, contextual-anchoring; zero-shot = 0.3898
Best estimation MAE1.86 SP @ qwen2.5:7bqwen2.5:3b = 2.91; zero-shot baseline = 5.7150
H2 baseline MAE2.03 SPpredictor constant = 4.04 SP (mean SP, not a MAE)
H1 Wilcoxonp ≈ 6.9×10⁻¹³, r = −0.51reject H0₁
H2 Wilcoxonp = 0.104, r = −0.16fail to reject H0₂
Executed sweep27 configs / 2,700 inferences9 models × 3 scenarios × 1 strategy, N=100
Energy / carbon≈68 Wh / ≈11.75 gCO₂eqfull 27-run sweep

Source: PUMA project documentation + reconciliation verdict · Confidence: verified-at-primary-source