An adversarial test we run on ourselves.

Study selection

Nine published gene-expression studies, 2004–2022, each with public data and a computational methods section, and none involving Genomarker. Round 1 deliberately mixes calibration cases — three papers later retracted, a signature that failed in phase III, a canonical confounding case — with highly cited studies and a clean counterexample.

Spec reconstruction

Each paper's methods are transcribed into a canonical analysis request exactly as published — ambiguities included. Where the publication under-specifies, the ambiguity itself becomes a finding, not a silent guess.

Re-execution

The reconstructed analysis runs on our production system, through the same checks every user's analysis gets. Nothing bespoke, nothing softened.

Classification

Findings are classified by severity — hard refusal, blocking, warning, confirmation — and by subject: the study, or Genomarker itself. A blocking finding stops the analysis until someone signs a justification; a hard refusal stops it outright.

Disclosure ethics

Aggregate results are public. Per-paper findings are shared with the original authors first, anonymized in public reporting, and never framed as accusations — the audit tests Genomarker, not the scientists.

Versioning

This is round 1 of a living audit, run on production in 2026. Round 2 re-runs the corpus through the replay engine when it ships, and every revision is archived.