Claim boundary

Supported evidence statements

• Twenty balanced five-fold realizations each cover all 2,160 targets exactly once out of fold.
• RMSE improvements versus zero and train mean are stable under joint realization/target uncertainty.
• Perturbation-specific residual cosine is stable above zero.
• Raw-cosine point estimates are positive in 20/20 realizations, but its joint 95% interval crosses zero; directional evidence is mixed.
• No negative control supports both RMSE baselines in any realization.
• Historical overlapping splits remain secondary sensitivity records.

Unsupported inferences

• Fold-robust raw directional prediction or strong gene-specific direction.
• Independent biological replication or cross-dataset/cross-cell-type generalization.
• Muscle or DMD perturbation-response validity.
• DMD mechanism, target validation, therapeutic efficacy, safety or clinical utility.
• A universal model leaderboard.

Evidence facts

The direct strict endpoint passed 0/16 outcomes while 16/16 legacy RMSE-comparator outcomes passed. The contrast is retained to show how endpoint and baseline definitions change the conclusion; it is not presented as evidence that all directional gates pass.

The current public website is deployed; DOI/archive deposition and manuscript submission remain pending.