Data v1 Fit Report¶

Archive. This page records an earlier frozen experiment snapshot. The final report and current Findings page are authoritative.

Archive note: this is a frozen fit-report snapshot kept for audit context. Current runnable outputs are under outputs/transformer/ and Chapter 4 evidence outputs are under outputs/transformer/supplemental/. Use python -m src.transformer_pipeline.run_all --run --train E31 for the current mainline.

policy: v116_gapfill_dual_goodorig_nm40_ms10_smc_s20260876
protocol path: ./outputs/external_benchmarks/e311_but_node_ladder_tuning_10s_2026_06_08/analysis/good_medium_geometry_repair/clean_but_protocols/v116_gapfill_dual_goodorig_nm40_ms10_smc_s20260876
split path: ./outputs/external_benchmarks/e311_but_node_ladder_tuning_10s_2026_06_08/analysis/good_medium_geometry_repair/event_factorized_sqi_conformer/rh_splits/v116_gapfill_dual_goodorig_nm40__k1_s20260876/fold0
selector runtime: CPU numpy/sklearn; --device auto is recorded but not used by the current data selector.

Protocol Counts¶

v116_candidate_type original_but but_native_morph ptb_morph clean_style class_name
good 10530 0 0 0 medium 6449 1616 2424 41 bad 1656 3514 5271 89

Protocol Percent By Class¶

v116_candidate_type original_but but_native_morph ptb_morph clean_style class_name
good 100.00 0.00 0.00 0.00 medium 61.24 15.35 23.02 0.39 bad 15.73 33.37 50.06 0.85

Generated Gap Composition¶

v116_candidate_type but_native_morph ptb_morph clean_style but_native_morph_pct ptb_morph_pct clean_style_pct class_name
medium 1616 2424 41 39.6 59.4 1.0 bad 3514 5271 89 39.6 59.4 1.0

Split Counts¶

class_name good medium bad split
train 8310 8310 8310 val 1053 618 178 test 1053 632 164

Train Composition¶

v116_candidate_type original_but but_native_morph ptb_morph clean_style class_name
good 8310 0 0 0 medium 5199 1061 2019 31 bad 1314 2393 4535 68

Train Percent By Class¶

v116_candidate_type original_but but_native_morph ptb_morph clean_style class_name
good 100.00 0.00 0.00 0.00 medium 62.56 12.77 24.30 0.37 bad 15.81 28.80 54.57 0.82

Pool Capacity And Selector Result¶

              class_name gap_fill_component  selected_n  pool_n   rbf_mmd  sym_domain_auc  pca_density_overlap  pool_to_selected

60 medium_clean_style clean_style 41 300 0.259555 1.000000 0.137337 7.32 145 medium_but_native_morph but_native_morph 1616 51592 0.006918 0.994700 0.810238 31.93 346 medium_ptb_morph ptb_morph 2424 6000 0.164117 1.000000 0.716124 2.48 407 bad_clean_style clean_style 89 300 0.482773 1.000000 0.004380 3.37 492 bad_but_native_morph but_native_morph 3514 13248 0.044235 0.993951 0.900477 3.77 693 bad_ptb_morph ptb_morph 5271 7000 0.134500 1.000000 0.643726 1.33

Distribution Metrics¶

      scope  but_n  synthetic_n   rbf_mmd  sliced_wasserstein  quantile_loss  sym_domain_auc  pca_density_overlap

0 all_labels 18635 31590 0.059592 0.346813 0.335463 0.704572 0.775698 1 class_good 10530 10530 0.001377 0.055613 0.000000 0.509960 1.000000 2 class_medium 6449 10530 0.008691 0.117459 0.036022 0.649138 0.924499 3 class_bad 1656 10530 0.045898 0.229435 0.199833 0.933622 0.798753

Raw Feature Alias Consistency¶

                  check  max_abs_delta

0 raw_rms_vs_rms 0.000003 1 raw_ptp_vs_ptp 0.000020 2 raw_diff_vs_non_qrs_diff 0.000045

Top Feature Drift Versus Original¶

class_name candidate_type rows top_drift 0 medium but_native_morph 1616 sqi_bSQI=0.543; raw_rms=0.028; raw_ptp_p99_p01=0.026; raw_diff_abs_p95=0.010; band_5_15=0.006 1 medium ptb_morph 2424 band_5_15=2.142; pca_margin=0.211; sqi_bSQI=0.156; knn_label_purity=0.132; raw_diff_abs_p95=0.083 2 medium clean_style 41 band_5_15=2.142; knn_label_purity=0.628; sqi_bSQI=0.474; band_15_30=0.468; raw_diff_abs_p95=0.353 3 bad but_native_morph 3514 sqi_bSQI=4.476; raw_rms=0.067; band_15_30=0.020; hf_ratio=0.020; raw_diff_abs_p95=0.011 4 bad ptb_morph 5271 band_5_15=5.359; pca_margin=4.611; sqi_bSQI=2.135; band_15_30=0.842; raw_rms=0.557 5 bad clean_style 89 band_15_30=15.861; sqi_bSQI=12.473; raw_diff_abs_p95=12.276; pca_margin=6.889; band_5_15=5.359

Gates¶

protocol rows: 31590
val/test generated rows: 0
final selected count is hard-gated by selector; insufficient pools raise instead of silently underfilling.

Figures¶

./docs/data_v1_figures/distribution_domain_auc.png
./docs/data_v1_figures/distribution_pca_overlap.png
./docs/data_v1_figures/distribution_rbf_mmd.png
./docs/data_v1_figures/distribution_sliced_wasserstein.png
./docs/data_v1_figures/generated_gap_component_share.png
./docs/data_v1_figures/original_but_class_counts.png
./docs/data_v1_figures/split_class_counts.png
./docs/data_v1_figures/train_candidate_type_composition.png