# H2 unique-target cross-fitted response-independent ridge audit

Verification Status: COMPUTED_PENDING_INDEPENDENT_RECOMPUTATION

Targets: 2160; output genes: 2000; safe features: 32; outer folds: 5.

The primary feature set contains only DMD-consensus and priority-v1 columns. Dataset-wide AnnData summaries and DepMap are excluded. Imputation, scaling and alpha selection are outer-training-fold local.

Negative controls reuse the primary fold-specific alpha. Any negative-control RMSE-gate pass invalidates the primary RMSE interpretation.

## Model summary

| model | n_unique_targets | mean_model_rmse | mean_zero_rmse | mean_train_mean_rmse | mean_model_cosine | mean_train_mean_cosine | mean_residual_cosine | relative_rmse_improvement_vs_zero | relative_rmse_improvement_vs_train_mean | raw_cosine_difference_vs_train_mean |
| --- | --- | --- | --- | --- | --- | --- | --- | --- | --- | --- |
| negative_feature_row_shuffle | 2160 | 0.118169 | 0.123799 | 0.117965 | 0.347089 | 0.348853 | -0.00621292 | 0.0454766 | -0.00172654 | -0.00176375 |
| negative_outcome_coordinate_permutation | 2160 | 0.139429 | 0.123799 | 0.117965 | -0.00174807 | 0.348853 | 0.0987155 | -0.126258 | -0.181954 | -0.350601 |
| negative_response_row_shuffle | 2160 | 0.118179 | 0.123799 | 0.117965 | 0.347212 | 0.348853 | -0.0069696 | 0.0453904 | -0.00181701 | -0.00164058 |
| safe_external_v1 | 2160 | 0.117485 | 0.123799 | 0.117965 | 0.35049 | 0.348853 | 0.0454424 | 0.0509996 | 0.00406958 | 0.00163656 |

## Frozen gates

| model | rmse_composite_gate_before_negative_control_check | raw_directional_gate | residual_directional_gate | strict_raw_composite_gate_before_negative_control_check | any_negative_control_rmse_gate_pass | final_rmse_composite_gate | final_strict_raw_composite_gate | negative_control_invalidation |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| negative_feature_row_shuffle | FAIL | FAIL | FAIL | FAIL | False | FAIL | FAIL | NOT_APPLICABLE |
| negative_outcome_coordinate_permutation | FAIL | FAIL | PASS | FAIL | False | FAIL | FAIL | NOT_APPLICABLE |
| negative_response_row_shuffle | FAIL | FAIL | FAIL | FAIL | False | FAIL | FAIL | NOT_APPLICABLE |
| safe_external_v1 | PASS | PASS | PASS | PASS | False | PASS | PASS | CLEAR |

Raw directional failure cannot be rescued by residual cosine. This HepG2 analysis is retrospective but response-independent with respect to held-out target outcomes.