# Feature-family ablation and leakage-sensitivity audit

The audit reuses the five frozen 25% held-out-gene splits. Every non-control feature model selects its ridge penalty on training genes only. Two shuffled controls test whether gene-feature assignment and feature-to-delta association are required.

## Gate reproducibility

| model | n_splits | paired_rmse_pass | residual_specific_pass | raw_composite_pass |
| --- | --- | --- | --- | --- |
| substrate_independent_all | 5 | 5 | 5 | 0 |
| v1_depmap_only | 5 | 5 | 5 | 0 |
| v1_full | 5 | 5 | 5 | 0 |
| v1_no_dataset_var | 5 | 5 | 5 | 0 |
| v1_no_dmd_consensus | 5 | 5 | 5 | 0 |
| v2_full | 5 | 5 | 5 | 0 |
| v2_no_organoid | 5 | 5 | 5 | 0 |
| v2_no_patient | 5 | 5 | 5 | 0 |
| v2_no_sparse_curated | 5 | 5 | 5 | 0 |
| v1_no_depmap | 5 | 4 | 4 | 1 |
| v1_var_only | 5 | 3 | 3 | 1 |
| v1_dmd_consensus_only | 5 | 1 | 0 | 0 |
| negative_feature_row_shuffle | 5 | 0 | 0 | 0 |
| negative_training_delta_shuffle | 5 | 0 | 0 | 0 |
| v2_added_only | 5 | 0 | 0 | 0 |

## Cross-seed aggregate

| model | n_splits | mean_model_rmse | mean_relative_improvement_vs_zero | mean_relative_improvement_vs_train_mean | mean_perturbation_specific_residual_cosine | mean_raw_cosine_improvement_vs_train_mean |
| --- | --- | --- | --- | --- | --- | --- |
| v1_no_dmd_consensus | 5 | 0.111693 | 0.09329 | 0.0502583 | 0.258507 | 0.000704356 |
| v1_full | 5 | 0.11198 | 0.0909476 | 0.0478095 | 0.255407 | -0.00177981 |
| v2_no_sparse_curated | 5 | 0.112531 | 0.0864829 | 0.0431304 | 0.234522 | 0.00187491 |
| v1_depmap_only | 5 | 0.112672 | 0.0853139 | 0.0419149 | 0.251056 | 0.000952901 |
| v1_no_dataset_var | 5 | 0.112717 | 0.0849328 | 0.0415221 | 0.255955 | -0.0011702 |
| v2_no_organoid | 5 | 0.112758 | 0.0846291 | 0.0411875 | 0.236373 | 0.00323313 |
| v2_full | 5 | 0.11281 | 0.0842098 | 0.0407482 | 0.23449 | 0.00194966 |
| v2_no_patient | 5 | 0.112833 | 0.0840166 | 0.0405467 | 0.2347 | 0.00173403 |
| substrate_independent_all | 5 | 0.11332 | 0.0800495 | 0.0363981 | 0.233382 | 0.00152134 |
| v1_var_only | 5 | 0.114864 | 0.067577 | 0.0233094 | 0.131199 | 0.00316035 |
| v1_no_depmap | 5 | 0.115337 | 0.063698 | 0.0192622 | 0.125548 | 0.0042509 |
| v1_dmd_consensus_only | 5 | 0.117106 | 0.0492808 | 0.00418224 | 0.0405419 | 0.00164429 |
| v2_added_only | 5 | 0.117195 | 0.0485656 | 0.00342009 | 0.0900777 | 0.000859248 |
| negative_training_delta_shuffle | 5 | 0.118636 | 0.0368887 | -0.00880546 | -0.00186222 | -0.00889349 |
| negative_feature_row_shuffle | 5 | 0.11873 | 0.0361175 | -0.0096141 | -0.00327308 | -0.0089943 |

## Ablation effects relative to the corresponding full model

| model | reference_model | comparison | n_splits | mean_difference | min_difference | max_difference | holm_significant_splits |
| --- | --- | --- | --- | --- | --- | --- | --- |
| negative_feature_row_shuffle | v1_full | residual_cosine_ablation_minus_full | 5 | -0.25868 | -0.294835 | -0.237357 | 5 |
| negative_training_delta_shuffle | v1_full | residual_cosine_ablation_minus_full | 5 | -0.257269 | -0.276156 | -0.24761 | 5 |
| v1_dmd_consensus_only | v1_full | residual_cosine_ablation_minus_full | 5 | -0.214865 | -0.236259 | -0.189455 | 5 |
| v1_no_depmap | v1_full | residual_cosine_ablation_minus_full | 5 | -0.129858 | -0.152869 | -0.114289 | 5 |
| v1_var_only | v1_full | residual_cosine_ablation_minus_full | 5 | -0.124208 | -0.143821 | -0.10896 | 5 |
| v1_depmap_only | v1_full | residual_cosine_ablation_minus_full | 5 | -0.00435087 | -0.00922305 | 0.00423273 | 0 |
| v1_no_dataset_var | v1_full | residual_cosine_ablation_minus_full | 5 | 0.000547946 | -0.00631363 | 0.00390386 | 0 |
| v1_no_dmd_consensus | v1_full | residual_cosine_ablation_minus_full | 5 | 0.00310035 | -0.000205603 | 0.00545772 | 0 |
| v1_no_dmd_consensus | v1_full | rmse_ablation_minus_full | 5 | -0.000286756 | -0.000455679 | -0.000149554 | 0 |
| v1_depmap_only | v1_full | rmse_ablation_minus_full | 5 | 0.000692139 | 0.000347136 | 0.00116059 | 0 |
| v1_no_dataset_var | v1_full | rmse_ablation_minus_full | 5 | 0.000737365 | 0.000420223 | 0.00132426 | 1 |
| v1_var_only | v1_full | rmse_ablation_minus_full | 5 | 0.00288418 | 0.00205525 | 0.00338809 | 4 |
| v1_no_depmap | v1_full | rmse_ablation_minus_full | 5 | 0.00335731 | 0.00305016 | 0.00374893 | 5 |
| v1_dmd_consensus_only | v1_full | rmse_ablation_minus_full | 5 | 0.00512624 | 0.00462265 | 0.00571334 | 5 |
| negative_training_delta_shuffle | v1_full | rmse_ablation_minus_full | 5 | 0.00665617 | 0.00622228 | 0.00701759 | 5 |
| negative_feature_row_shuffle | v1_full | rmse_ablation_minus_full | 5 | 0.0067504 | 0.00620302 | 0.00728452 | 5 |
| v2_added_only | v2_full | residual_cosine_ablation_minus_full | 5 | -0.144412 | -0.162877 | -0.126311 | 5 |
| substrate_independent_all | v2_full | residual_cosine_ablation_minus_full | 5 | -0.00110816 | -0.00656015 | 0.00456278 | 0 |
| v2_no_sparse_curated | v2_full | residual_cosine_ablation_minus_full | 5 | 3.19134e-05 | -0.000704674 | 0.000857583 | 0 |
| v2_no_patient | v2_full | residual_cosine_ablation_minus_full | 5 | 0.00020977 | -0.0016168 | 0.00142435 | 0 |
| v2_no_organoid | v2_full | residual_cosine_ablation_minus_full | 5 | 0.00188319 | -0.000455907 | 0.0063065 | 0 |
| v2_no_sparse_curated | v2_full | rmse_ablation_minus_full | 5 | -0.000278397 | -0.000870175 | -4.1158e-05 | 0 |
| v2_no_organoid | v2_full | rmse_ablation_minus_full | 5 | -5.18867e-05 | -0.000126165 | 6.3808e-07 | 0 |
| v2_no_patient | v2_full | rmse_ablation_minus_full | 5 | 2.35958e-05 | -1.28884e-05 | 4.72212e-05 | 0 |
| substrate_independent_all | v2_full | rmse_ablation_minus_full | 5 | 0.000510154 | 0.000144284 | 0.000867638 | 1 |
| v2_added_only | v2_full | rmse_ablation_minus_full | 5 | 0.00438561 | 0.00384144 | 0.00476783 | 5 |

## Per-split effects

| model | seed | n_test_genes | mean_model_rmse | mean_zero_rmse | mean_train_mean_rmse | mean_model_cosine | mean_train_mean_cosine | mean_perturbation_specific_residual_cosine | relative_rmse_improvement_vs_zero | relative_rmse_improvement_vs_train_mean | raw_cosine_difference_vs_train_mean |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| negative_feature_row_shuffle | 20260712 | 540 | 0.117484 | 0.122424 | 0.11662 | 0.341897 | 0.350142 | 0.0141892 | 0.0403522 | -0.00741099 | -0.00824545 |
| negative_feature_row_shuffle | 20260713 | 540 | 0.120564 | 0.125713 | 0.119704 | 0.342671 | 0.350854 | 0.00513456 | 0.0409561 | -0.00718804 | -0.00818339 |
| negative_feature_row_shuffle | 20260714 | 540 | 0.119146 | 0.123006 | 0.117724 | 0.332697 | 0.341558 | -0.0305315 | 0.0313842 | -0.0120788 | -0.00886056 |
| negative_feature_row_shuffle | 20260715 | 540 | 0.116587 | 0.119611 | 0.11551 | 0.325522 | 0.332757 | 0.0113388 | 0.025288 | -0.00932341 | -0.00723582 |
| negative_feature_row_shuffle | 20260716 | 540 | 0.11987 | 0.125205 | 0.11844 | 0.346037 | 0.358483 | -0.0164965 | 0.0426068 | -0.0120692 | -0.0124463 |
| negative_training_delta_shuffle | 20260712 | 540 | 0.117863 | 0.122424 | 0.11662 | 0.339024 | 0.350142 | -0.021143 | 0.0372587 | -0.0106584 | -0.0111182 |
| negative_training_delta_shuffle | 20260713 | 540 | 0.121189 | 0.125713 | 0.119704 | 0.341453 | 0.350854 | -0.0108984 | 0.035989 | -0.0124045 | -0.00940111 |
| negative_training_delta_shuffle | 20260714 | 540 | 0.118432 | 0.123006 | 0.117724 | 0.334121 | 0.341558 | 0.0109639 | 0.0371862 | -0.00601646 | -0.00743729 |
| negative_training_delta_shuffle | 20260715 | 540 | 0.116479 | 0.119611 | 0.11551 | 0.3225 | 0.332757 | 0.01104 | 0.0261849 | -0.00839473 | -0.0102574 |
| negative_training_delta_shuffle | 20260716 | 540 | 0.119217 | 0.125205 | 0.11844 | 0.35223 | 0.358483 | 0.000726382 | 0.0478248 | -0.00655321 | -0.00625343 |
| substrate_independent_all | 20260712 | 540 | 0.113022 | 0.122424 | 0.11662 | 0.351316 | 0.350142 | 0.236102 | 0.076796 | 0.0308468 | 0.00117414 |
| substrate_independent_all | 20260713 | 540 | 0.115661 | 0.125713 | 0.119704 | 0.350069 | 0.350854 | 0.224009 | 0.0799604 | 0.0337743 | -0.000785017 |
| substrate_independent_all | 20260714 | 540 | 0.113199 | 0.123006 | 0.117724 | 0.342648 | 0.341558 | 0.231754 | 0.0797313 | 0.0384377 | 0.00109025 |
| substrate_independent_all | 20260715 | 540 | 0.110967 | 0.119611 | 0.11551 | 0.336964 | 0.332757 | 0.240424 | 0.072269 | 0.0393258 | 0.00420696 |
| substrate_independent_all | 20260716 | 540 | 0.11375 | 0.125205 | 0.11844 | 0.360403 | 0.358483 | 0.234621 | 0.0914905 | 0.0396062 | 0.00192039 |
| v1_depmap_only | 20260712 | 540 | 0.111628 | 0.122424 | 0.11662 | 0.35348 | 0.350142 | 0.259246 | 0.0881897 | 0.0428075 | 0.00333751 |
| v1_depmap_only | 20260713 | 540 | 0.115032 | 0.125713 | 0.119704 | 0.350926 | 0.350854 | 0.236663 | 0.084961 | 0.0390259 | 7.20909e-05 |
| v1_depmap_only | 20260714 | 540 | 0.112583 | 0.123006 | 0.117724 | 0.344873 | 0.341558 | 0.259913 | 0.0847392 | 0.0436703 | 0.00331548 |
| v1_depmap_only | 20260715 | 540 | 0.110622 | 0.119611 | 0.11551 | 0.334475 | 0.332757 | 0.257667 | 0.0751518 | 0.042311 | 0.0017176 |
| v1_depmap_only | 20260716 | 540 | 0.113494 | 0.125205 | 0.11844 | 0.354805 | 0.358483 | 0.241791 | 0.0935279 | 0.0417599 | -0.00367817 |
| v1_dmd_consensus_only | 20260712 | 540 | 0.116239 | 0.122424 | 0.11662 | 0.352137 | 0.350142 | 0.0365858 | 0.0505237 | 0.00326676 | 0.00199475 |
| v1_dmd_consensus_only | 20260713 | 540 | 0.118984 | 0.125713 | 0.119704 | 0.352732 | 0.350854 | 0.0530362 | 0.0535274 | 0.00601433 | 0.00187746 |
| v1_dmd_consensus_only | 20260714 | 540 | 0.117365 | 0.123006 | 0.117724 | 0.343816 | 0.341558 | 0.0291989 | 0.0458658 | 0.00305251 | 0.00225816 |
| v1_dmd_consensus_only | 20260715 | 540 | 0.115175 | 0.119611 | 0.11551 | 0.332043 | 0.332757 | 0.0306305 | 0.0370889 | 0.00289655 | -0.00071441 |
| v1_dmd_consensus_only | 20260716 | 540 | 0.117768 | 0.125205 | 0.11844 | 0.361289 | 0.358483 | 0.0532583 | 0.0593981 | 0.00568102 | 0.0028055 |
| v1_full | 20260712 | 540 | 0.11128 | 0.122424 | 0.11662 | 0.350793 | 0.350142 | 0.255013 | 0.0910252 | 0.0457842 | 0.000651071 |
| v1_full | 20260713 | 540 | 0.114361 | 0.125713 | 0.119704 | 0.347704 | 0.350854 | 0.242491 | 0.0902989 | 0.0446317 | -0.00315013 |
| v1_full | 20260714 | 540 | 0.11221 | 0.123006 | 0.117724 | 0.340734 | 0.341558 | 0.264303 | 0.0877713 | 0.0468384 | -0.000823819 |
| v1_full | 20260715 | 540 | 0.109462 | 0.119611 | 0.11551 | 0.331329 | 0.332757 | 0.26689 | 0.0848548 | 0.0523585 | -0.00142843 |
| v1_full | 20260716 | 540 | 0.112585 | 0.125205 | 0.11844 | 0.354335 | 0.358483 | 0.248336 | 0.100788 | 0.0494344 | -0.00414776 |
| v1_no_dataset_var | 20260712 | 540 | 0.111815 | 0.122424 | 0.11662 | 0.352128 | 0.350142 | 0.258917 | 0.0866554 | 0.0411969 | 0.00198542 |
| v1_no_dataset_var | 20260713 | 540 | 0.115113 | 0.125713 | 0.119704 | 0.347835 | 0.350854 | 0.24508 | 0.084318 | 0.0383506 | -0.00301961 |
| v1_no_dataset_var | 20260714 | 540 | 0.11263 | 0.123006 | 0.117724 | 0.341077 | 0.341558 | 0.265242 | 0.084355 | 0.0432688 | -0.000480728 |
| v1_no_dataset_var | 20260715 | 540 | 0.110786 | 0.119611 | 0.11551 | 0.332262 | 0.332757 | 0.260576 | 0.0737835 | 0.0408941 | -0.000495197 |
| v1_no_dataset_var | 20260716 | 540 | 0.113241 | 0.125205 | 0.11844 | 0.354642 | 0.358483 | 0.249958 | 0.0955523 | 0.0439 | -0.0038409 |
| v1_no_depmap | 20260712 | 540 | 0.114473 | 0.122424 | 0.11662 | 0.354556 | 0.350142 | 0.120784 | 0.0649503 | 0.0184114 | 0.00441357 |
| v1_no_depmap | 20260713 | 540 | 0.117411 | 0.125713 | 0.119704 | 0.35506 | 0.350854 | 0.128203 | 0.066036 | 0.0191509 | 0.00420524 |
| v1_no_depmap | 20260714 | 540 | 0.115959 | 0.123006 | 0.117724 | 0.346078 | 0.341558 | 0.111434 | 0.0572937 | 0.0149933 | 0.00452042 |
| v1_no_depmap | 20260715 | 540 | 0.112566 | 0.119611 | 0.11551 | 0.33507 | 0.332757 | 0.147703 | 0.0589047 | 0.025487 | 0.00231307 |
| v1_no_depmap | 20260716 | 540 | 0.116277 | 0.125205 | 0.11844 | 0.364285 | 0.358483 | 0.119618 | 0.0713054 | 0.0182683 | 0.0058022 |
| v1_no_dmd_consensus | 20260712 | 540 | 0.110858 | 0.122424 | 0.11662 | 0.350769 | 0.350142 | 0.260471 | 0.0944778 | 0.0494085 | 0.000626431 |
| v1_no_dmd_consensus | 20260713 | 540 | 0.114108 | 0.125713 | 0.119704 | 0.350823 | 0.350854 | 0.242286 | 0.0923135 | 0.0467475 | -3.13483e-05 |
| v1_no_dmd_consensus | 20260714 | 540 | 0.112057 | 0.123006 | 0.117724 | 0.343863 | 0.341558 | 0.267409 | 0.0890119 | 0.0481347 | 0.00230504 |
| v1_no_dmd_consensus | 20260715 | 540 | 0.109006 | 0.119611 | 0.11551 | 0.336381 | 0.332757 | 0.2705 | 0.0886644 | 0.0563035 | 0.00362346 |
| v1_no_dmd_consensus | 20260716 | 540 | 0.112436 | 0.125205 | 0.11844 | 0.355481 | 0.358483 | 0.251869 | 0.101982 | 0.0506971 | -0.0030018 |
| v1_var_only | 20260712 | 540 | 0.114174 | 0.122424 | 0.11662 | 0.351724 | 0.350142 | 0.132954 | 0.0673899 | 0.0209724 | 0.00158197 |
| v1_var_only | 20260713 | 540 | 0.11714 | 0.125713 | 0.119704 | 0.353106 | 0.350854 | 0.124554 | 0.0681941 | 0.0214173 | 0.00225162 |
| v1_var_only | 20260714 | 540 | 0.115515 | 0.123006 | 0.117724 | 0.344402 | 0.341558 | 0.120482 | 0.0609014 | 0.0187628 | 0.00284379 |
| v1_var_only | 20260715 | 540 | 0.111517 | 0.119611 | 0.11551 | 0.337778 | 0.332757 | 0.15793 | 0.0676721 | 0.0345657 | 0.00502075 |
| v1_var_only | 20260716 | 540 | 0.115974 | 0.125205 | 0.11844 | 0.362587 | 0.358483 | 0.120075 | 0.0737273 | 0.0208286 | 0.00410361 |
| v2_added_only | 20260712 | 540 | 0.11707 | 0.122424 | 0.11662 | 0.349935 | 0.350142 | 0.078771 | 0.0437354 | -0.00385934 | -0.000207356 |
| v2_added_only | 20260713 | 540 | 0.118879 | 0.125713 | 0.119704 | 0.352649 | 0.350854 | 0.0914549 | 0.0543645 | 0.00689342 | 0.00179423 |
| v2_added_only | 20260714 | 540 | 0.117174 | 0.123006 | 0.117724 | 0.34291 | 0.341558 | 0.0812529 | 0.0474141 | 0.00467029 | 0.00135194 |
| v2_added_only | 20260715 | 540 | 0.114867 | 0.119611 | 0.11551 | 0.334212 | 0.332757 | 0.095163 | 0.0396618 | 0.0055608 | 0.00145456 |
| v2_added_only | 20260716 | 540 | 0.117986 | 0.125205 | 0.11844 | 0.358386 | 0.358483 | 0.103747 | 0.0576521 | 0.00383529 | -9.71281e-05 |
| v2_full | 20260712 | 540 | 0.112445 | 0.122424 | 0.11662 | 0.35227 | 0.350142 | 0.241648 | 0.0815094 | 0.0357947 | 0.00212732 |
| v2_full | 20260713 | 540 | 0.115037 | 0.125713 | 0.119704 | 0.351307 | 0.350854 | 0.224254 | 0.0849217 | 0.0389847 | 0.000452759 |
| v2_full | 20260714 | 540 | 0.113055 | 0.123006 | 0.117724 | 0.342917 | 0.341558 | 0.229508 | 0.0809043 | 0.0396633 | 0.00135936 |
| v2_full | 20260715 | 540 | 0.1101 | 0.119611 | 0.11551 | 0.336984 | 0.332757 | 0.246984 | 0.0795228 | 0.0468372 | 0.00422611 |
| v2_full | 20260716 | 540 | 0.113411 | 0.125205 | 0.11844 | 0.360066 | 0.358483 | 0.230058 | 0.094191 | 0.0424609 | 0.00158278 |
| v2_no_organoid | 20260712 | 540 | 0.112446 | 0.122424 | 0.11662 | 0.353068 | 0.350142 | 0.241192 | 0.0815042 | 0.0357892 | 0.00292598 |
| v2_no_organoid | 20260713 | 540 | 0.114989 | 0.125713 | 0.119704 | 0.352043 | 0.350854 | 0.224211 | 0.0853012 | 0.0393831 | 0.00118882 |
| v2_no_organoid | 20260714 | 540 | 0.112928 | 0.123006 | 0.117724 | 0.34582 | 0.341558 | 0.235814 | 0.08193 | 0.040735 | 0.00426245 |
| v2_no_organoid | 20260715 | 540 | 0.110079 | 0.119611 | 0.11551 | 0.337863 | 0.332757 | 0.247918 | 0.0796926 | 0.0470131 | 0.00510542 |
| v2_no_organoid | 20260716 | 540 | 0.113346 | 0.125205 | 0.11844 | 0.361166 | 0.358483 | 0.232733 | 0.0947173 | 0.0430172 | 0.002683 |
| v2_no_patient | 20260712 | 540 | 0.112476 | 0.122424 | 0.11662 | 0.352756 | 0.350142 | 0.240031 | 0.0812597 | 0.0355326 | 0.00261364 |
| v2_no_patient | 20260713 | 540 | 0.115077 | 0.125713 | 0.119704 | 0.351031 | 0.350854 | 0.22513 | 0.0846074 | 0.0386546 | 0.000176704 |
| v2_no_patient | 20260714 | 540 | 0.113068 | 0.123006 | 0.117724 | 0.342127 | 0.341558 | 0.229495 | 0.080794 | 0.039548 | 0.000569101 |
| v2_no_patient | 20260715 | 540 | 0.110147 | 0.119611 | 0.11551 | 0.336338 | 0.332757 | 0.247361 | 0.079128 | 0.0464284 | 0.00358008 |
| v2_no_patient | 20260716 | 540 | 0.113399 | 0.125205 | 0.11844 | 0.360214 | 0.358483 | 0.231482 | 0.0942939 | 0.0425697 | 0.00173064 |
| v2_no_sparse_curated | 20260712 | 540 | 0.111575 | 0.122424 | 0.11662 | 0.352495 | 0.350142 | 0.241031 | 0.0886173 | 0.0432563 | 0.00235232 |
| v2_no_sparse_curated | 20260713 | 540 | 0.114996 | 0.125713 | 0.119704 | 0.35135 | 0.350854 | 0.225001 | 0.0852491 | 0.0393285 | 0.000495258 |
| v2_no_sparse_curated | 20260714 | 540 | 0.112905 | 0.123006 | 0.117724 | 0.34225 | 0.341558 | 0.230365 | 0.0821174 | 0.0409308 | 0.000692261 |
| v2_no_sparse_curated | 20260715 | 540 | 0.109913 | 0.119611 | 0.11551 | 0.336716 | 0.332757 | 0.246279 | 0.0810798 | 0.0484495 | 0.00395868 |
| v2_no_sparse_curated | 20260716 | 540 | 0.113266 | 0.125205 | 0.11844 | 0.360359 | 0.358483 | 0.229935 | 0.0953507 | 0.0436868 | 0.00187604 |

## Leakage boundary

`dataset_var` features are global AnnData gene statistics and are conservatively marked as carrying nonzero substrate-summary leakage risk because they were not recomputed within each outer training fold. DepMap, DMD disease signatures, patient/organoid deltas, and curated disease metadata do not use the held-out HepG2 perturbation response. The `v1_no_dataset_var` and `substrate_independent_all` models test whether generalization survives removal of that potentially contaminated block.

## Interpretation rule

An ablation is considered informative only when effect size, bootstrap interval, paired randomization, multiplicity correction, and consistency across seeds agree. A shuffled control passing the same gate would invalidate the feature-specific interpretation.

## Statistical fallacy scan (11/11 checked)

- Simpson's paradox: no direction reversal was hidden by pooling because every gate and comparison is reported per outer split before cross-seed aggregation.
- Ecological fallacy: conclusions are restricted to perturbation-level prediction; no cell- or patient-level individual inference is made.
- Berkson's paradox: eligible perturbations are a filtered HepG2 CRISPRi subset, so transport beyond this substrate remains a stated limitation.
- Collider bias: no post-response covariate is conditioned on; feature selection is source-defined rather than outcome-filtered.
- Base-rate neglect: diagnostic prevalence metrics are not used; not applicable.
- Regression to the mean: there is no extreme-score enrolment or pre/post comparison; not applicable.
- Survivorship bias: exclusions are determined by fixed cell-count and feature-availability rules and are reported in the split manifest; conclusions do not cover excluded perturbations.
- Look-elsewhere effect: all prespecified feature families and both negative controls are reported, with Holm correction across the ablation family.
- Garden of forking paths: five outer splits are frozen and ridge penalties are selected only by inner training-fold CV; this remains an exploratory robustness audit rather than a preregistered confirmatory study.
- Correlation is not causation: feature attribution is predictive, not evidence that DepMap dependencies causally determine HepG2 transcriptional responses.
- Reverse causality: causal direction is not claimed; external feature-response association is interpreted only as predictive transfer.