Batch efficiency and model audit

Generated from measured MACSE and TreeBest members at 2026-09-19 02:22:50 UTC.

Advisory only: candidate models are saved for inspection but are never installed automatically. A recommendation to update should be reviewed across representative runs before replacing a frozen production model.

Current-run batching efficiency

stagemembers_plannedmembers_measuredsubmitted_batchesavoided_submissionssubmission_reduction_pctmedian_items_per_batchobserved_to_predicted_ratiorequested_walltime_utilization_pctpeak_memory_request_coverage_pct
MACSE21683216678752080895.96458 61.011206054.3820696.99769
TreeBest2213622136 232211399.896109640.978861727.4220586.95652

Submission reduction compares the number of orthogroups with the number of scheduler tasks. Requested-walltime utilization is measured member work divided by the sum of requested batch walltimes; it is a reservation-efficiency indicator, not CPU utilization.

Predicted versus observed batch duration

Frozen-model accuracy

Runtime prediction accuracy MACSE peak memory prediction accuracy

Candidate refits and recommendation

labelcurrent_model_versionobservationscurrent_maecandidate_cv_maecandidate_mae_improvement_pctcurrent_median_absolute_error_pctcurrent_observed_to_predicted_ratiorecommendationreason
MACSE runtimesalmonid-macse-runtime-gam-v3-2026-09-1421667145.073577148.99088-2.70021816.559091.0010251RETAINCurrent-run error and bias do not justify replacing the frozen model.
MACSE peak memorysalmonid-macse-memory-gam-v3-2026-09-1421667360.274065107.9155970.04625053.452170.6625558UPDATECurrent median error is 53.5% with 0.66x observed/predicted bias; the candidate reduces CV MAE by 70.0%.
TreeBest runtimesalmonid-treebest-runtime-gamma-gam-v2-2026-09-1422136 8.940376 7.0867620.73309431.318410.8258791UPDATECurrent median error is 31.3% with 0.83x observed/predicted bias; the candidate reduces CV MAE by 20.7%.
Candidate model cross-validation comparison
Decision rules: at least 30 valid observations; UPDATE requires current median absolute error of at least 25% or absolute bias of at least 20%, together with at least 15% lower cross-validated MAE. Unsafe memory coverage is also flagged. Otherwise the result is RETAIN, REVIEW, or INSUFFICIENT_DATA.

Machine-readable outputs