Repository navigation
Verified research-artifact recovery and reproduction, anchored to tested code commit 44918ef on the separate, unmerged reproduction PR branch. Main and all manuscript/supplement/figures/tables/historical results remain unchanged.
Contains three explicitly RETRAINED_NOT_ORIGINAL primary ResNet18-head replacements and 96 recovered ResNet18-based best states, with 48 additional resume files, configurations/scalers, per-model indexes, derived numerical exports, SHA256 manifests and reproducibility instructions. Original primary checkpoint identities remain missing; original doctor/infer do not silently accept replacements. No raw patient data or pretrained encoder is included. Artifact-specific CC-BY-NC-SA 4.0 terms and attributions are in LICENSE_AUDIT.md and the archives.
Verification: 22,500 deterministic runs and 3.375 million primary per-gene values pass strict 1e-10 archive checks; the new primary ensemble satisfies all advance-declared agreement criteria; the selected 22-patient GCN/local real-weight replay passes 792 full-precision reference rows. All 168 tests and 29 publication targets /71 checks pass. Scoped partial cases and retained runtime sensitivity are documented, not hidden.
Thirty-six verified Phikon heads and their derivative archives remain local pending custom-license clarification; the separate Phikon review draft contains metadata only. This public release does not grant rights to those withheld artifacts.
Download all assets, verify SHA256SUMS.txt, and follow README_RELEASE.md. RELEASE_VERIFICATION.json records the actual GitHub download/clean-checkout/real-image checks. The release tag may precede subsequent PR documentation receipts; its exact code commit is in the model/results manifest. The PR remains unmerged for author review.
Draft code/documentation PR #4 remains unmerged. Actual draft download/restore checks passed for all14 assets (706,622,974 bytes); all147 PT files loaded safely, all96 recovered public heads ran on a real554-location section, and the three replacement heads ran on eight real HEST patches with a fresh official encoder download. Cached and fresh encoder features matched exactly. The clean checkout reused the pinned dependency environment on the same machine; this is not a fresh-hardware replication. See RELEASE_VERIFICATION.json for exact scope and hashes.