Risk increased with candidate-minus-identity loss and selective area was negative.
Support-aware perturbation transfer
Can prediction failure be anticipated before the target outcome is observed?
This status page reports a completed GSE291147 historical replay and the retained GSE306429 primary-scorer failure. It keeps transport, computation, target scoring, custody and disease evidence separate.
Platform data remains live
This status page sits beside—not in place of—the Gene Evidence Atlas.
The status page does not remove or replace platform data. Open the released 17,921-gene corpus, inspect stable evidence records, or move directly to typed API objects.
Completed evidence
The GSE291147 direction signal did not produce candidate or decision benefit.
Across 139 items, rho was 0.182363 and selective-risk area was -0.002664. SOURCE_RIDGE loss was 0.917003 versus 0.914852 for identity. Hard support was 2/139; 35% requested coverage achieved 1.44% and decision loss was 0.915605. CPL01 was null and non-triviality failed.
Mean SOURCE_RIDGE loss exceeded the identity reference by 0.002152.
All hard-gated policies selected the same two supported items and remained above identity.
The strict independent-study increment remains zero.
Fixed historical replay update
Transport and plate statistics completed; the chemical replay stopped at its primary scorer.
GSE291147 scored all 139 items once. For GSE306429, all 52 authorized archives and 44 fixed plate sufficient-statistics outputs completed. The scorer then failed before target effects or scores were created. The failure is retained and the independent-group increment remains zero.
The audited from-zero repair completed at 2026-09-10 23:13:31 UTC.
Plate workers read target expression, but these intermediate outputs are not target effects or scores.
ValueError: Object arrays cannot be loaded when allow_pickle=False. No target effect, target score or Figure S3 was created.
No source refit, gene-panel change, route rescue, item rescue or substitute model is allowed.
Direction passed; mean candidate loss, decision loss, CPL01 and non-triviality did not support deployment.
Official GEO record ↗No A375 source checkpoint or target score was produced, and no substitute model was introduced.
The replay does not add an independently custodied study-level group.
The status records what completed, the exact failure boundary and what scientific objects do not exist.
Disease boundary
DMD is a motivation and evidence-triage use case, not an efficacy result.
Directly measured outcomes remain 0/21 for the current DMD candidate set. Restoring DMD as a central efficacy claim requires a separate disease-matched functional validation program.