results/benchmark_fixes/report.md

benchmark-triage-strategies

EXECUTIVE DECISION

• decision: USE cluster_only_triage
• reason: Model does not beat simple clustering in the independent benchmark.
• recommended_strategy_best_hit: random_cost_balanced
• recommended_strategy_best_tradeoff: reference_free_active_learning_v2::cluster_bandit
• recommended_strategy_safest: cluster_only_triage
• recommended_strategy_final: cluster_only_triage

Dataset validity

• dataset_dir: data/datasets/rdock_3PTB_25k_pubchem_local
• scientifically_valid_model_benchmark: True

Reference Ligand Baseline

• reference_ligand_available: True
• reference_ligand_baseline_score_n50: -8.56014

Benchmark mode

• independent_strategy_sampling: true
• shared_reference_stage: false
• reference_sample_docked_once_for_evaluation: True
• reference_sample_size: 300

Strategy comparison table

• cluster_only_triage: median_best_filtered_score -40.17, median_top5_mean_filtered_score -33.229092548674544, median_top10_mean_filtered_score -31.733016274337267, median_reduction_fraction 0.9492, median_total_runs_spent 124670.0, median_top5pct_recall 1.0
• cheap_descriptor_filter_only: median_best_filtered_score -40.1748, median_top5_mean_filtered_score -32.80636800080102, median_top10_mean_filtered_score -30.973934000400515, median_reduction_fraction 0.95, median_total_runs_spent 124560.0, median_top5pct_recall 0.7333333333333333
• random_cost_balanced: median_best_filtered_score -50.1126, median_top5_mean_filtered_score -36.376689471691215, median_top10_mean_filtered_score -33.623753936310976, median_reduction_fraction 0.9, median_total_runs_spent 125000.0, median_top5pct_recall 0.0
• diverse_random_cost_balanced: median_best_filtered_score -42.5011, median_top5_mean_filtered_score -35.63192, median_top10_mean_filtered_score -33.28833611220234, median_reduction_fraction 0.9, median_total_runs_spent 125000.0, median_top5pct_recall 0.26666666666666666
• single_fidelity_cost_balanced: median_best_filtered_score -42.8697, median_top5_mean_filtered_score -35.794622588693315, median_top10_mean_filtered_score -32.77415209590938, median_reduction_fraction 0.9, median_total_runs_spent 125000.0, median_top5pct_recall 0.3333333333333333
• reference_free_active_learning_v2::cluster_bandit: median_best_filtered_score -37.1844, median_top5_mean_filtered_score -34.26819999999999, median_top10_mean_filtered_score -32.90156, median_reduction_fraction 0.95, median_total_runs_spent 35345.0, median_top5pct_recall 0.8666666666666667

Plots

• results/benchmark_fixes/plots/strategy_comparison_recall_cost.png
• results/benchmark_fixes/plots/best_score_by_strategy.png
• results/benchmark_fixes/plots/false_negative_rate_by_strategy.png
• results/benchmark_fixes/plots/energy_timeline_per_strategy.png
• results/benchmark_fixes/plots/energy_timeline_per_strategy_per_seed.png
• results/benchmark_fixes/plots/energy_timeline_per_model_policy.png
• results/benchmark_fixes/plots/cumulative_best_filtered_score_vs_runs.png
• results/benchmark_fixes/plots/cumulative_best_filtered_score_vs_walltime.png
• results/benchmark_fixes/plots/topk_score_progression.png
• results/benchmark_fixes/plots/survivor_pool_size_timeline.png
• results/benchmark_fixes/plots/acquisition_component_timeline.png
• results/benchmark_fixes/plots/cluster_coverage_timeline.png
• results/benchmark_fixes/plots/promotion_funnel_per_strategy.png
• results/benchmark_fixes/plots/score_distribution_by_fidelity_and_strategy.png
• results/benchmark_fixes/plots/top_hits_component_breakdown.png
• results/benchmark_fixes/plots/top5_mean_score_by_strategy.png
• results/benchmark_fixes/plots/top10_mean_score_by_strategy.png
• results/benchmark_fixes/plots/reduction_vs_best_score_scatter.png
• results/benchmark_fixes/plots/cost_saved_vs_best_score_scatter.png
• results/benchmark_fixes/plots/cost_saved_vs_top10_score_scatter.png