Veriformatics: selection inside a global randomization test

A synthetic fixed panel: 8 experimental units, 35 binary features,
4 treated units and all 70 equally likely assignments.
Frozen selection gives 2/70; repeated selection gives 70/70.
These are exact constructed-example frequencies, not biological performance.

The fixed panel and selection algorithm must precede or remain invariant to
the observed assignment. The claim concerns a global sharp null, not a
selected-gene effect, partial-null FDR control, or biological validity.

Source identities: source/manifest.json retains original commits and hashes.
Local filesystem paths are intentionally omitted from this public manifest.
All 18 referenced evidence files preserve their original bytes. The exact
matrix and retained R tables are checked by independent Python enumeration:
2,450 scores and 140 p-value rows. The browser replays the resulting data.
It does not run Lean, sample assignments, accept visitor data, or verify the
full Python/R pipeline. This is a selective evidence export, not a complete
runnable research repository. No upstream runtime or package is bundled.

The original offline demo revision is 1c96770d3c6e7f34fd5dd779c1f820fbd7babc65.
The complete build and checks are retained in the website source project.
Original contributions: Copyright 2026 Veriformatics, Inc.
License: Apache-2.0 WITH LLVM-exception; see LICENSE and NOTICE.