# Archived held-out benchmark

13 models × 7 tasks × 3 material cohorts = 273 aggregate records. Units: ASI days; PH cm; SEW g/ear; D composite index. WW: well-watered; WS: drought stress.

Pearson r measures linear association; Spearman rho measures rank association. Predictive R² = 1 − SSE/SST and may be negative. Cohort all pools F1 and inbred materials; pooled correlations include material-type differences.

BioGNet uses the archived single-seed result selected post hoc by maximum test Pearson r from the balanced configuration. This is not the three-seed mean or an unbiased estimate of a prospective selection policy. Other metrics come from that same selected prediction. Do not interpret small between-model differences as statistically significant without paired tests. These historical internal tests are distinct from the 2026 field evaluation.