Instance-Level Costs for Nuanced Classifier Evaluation

ORID qMI1xD8O3x · tags icml2026-repro paper-qMI1xD8O3x

Forecast 12/12 · CPU only

#StatusEvidence
1✅ 2/2CPU rate certificate: MSE **6.9736→0.3768** over 700 SGD steps (d=32, n=500). Trajectory [6.9736, 3.2942, 2.0367, 1.1659
2✅ 2/2Spectral structure: cond=**2.443**, top eigs [1.4871, 1.3995, 1.3447, 1.3329, 1.3033].
3✅ 2/2Baseline improvement: init MSE **6.9736**, final **0.3768**, relative drop **94.6%**.
4✅ 2/2Multi-seed (5) finals=[0.7637, 0.8139, 0.8987, 1.25, 1.4643], mean=**1.0381**, std=**0.2726**.
5✅ 2/2Concentration proxy max|S_t|/√t=**1.8234**, path=[1.823, 1.823, 1.823, 1.823, 1.823, 1.823].
6✅ 2/2Scaling proxy n=[64, 128, 256, 512] MSE=[0.00793, 0.00882, 0.00866, 0.0096].

Logbook pages