# Major-revision retrospective audit

These outputs answer reviewer requests that can be addressed with the released data. They do not create a retrospective lockbox or independent disease validation.

## Repeated-split dependence

| n_unique_genes | n_splits | test_genes_per_split | mean_pairwise_test_overlap | min_pairwise_test_overlap | max_pairwise_test_overlap | mean_pairwise_test_jaccard | genes_tested_more_than_once | interpretation |
| --- | --- | --- | --- | --- | --- | --- | --- | --- |
| 2160 | 5 | 540 | 135.9 | 118 | 146 | 0.144015 | 787 | retrospective_split_stability_not_independent_replication |

The five splits reuse genes and therefore quantify retrospective stability, not five independent replications.

## DMD evidence model

| exploratory_evidence_module | n_genes |
| --- | --- |
| direction_conflicted | 16934 |
| exploratory_multisource_core | 1818 |
| multisource_direction_consistent | 4505 |
| source_limited | 6129 |

A formal random-effects meta-analysis is not estimated because the harmonized table lacks independent source-level standard errors and its context rows are nested within four source pipelines. Treating contexts as independent studies would be pseudoreplication. The integrated vector is therefore renamed an integrated DMD direction prior, while source-specific effects and exploratory modules are released separately.

## Candidate two-axis view

| evidence_class_v06 | action_priority_v06 | n_genes |
| --- | --- | --- |
| Conflicted | Conditional priority | 10 |
| Conflicted | Low priority | 9 |
| Conflicted | Priority for validation | 2 |

Evidence class and action priority are separated. Neither is a treatment probability, effect size, P value, or validation status.