Skip to content

fix: enforce the public source artifact boundary#225

Merged
abrichr merged 3 commits into
mainfrom
fix/source-boundary-carve-20260721
Jul 21, 2026
Merged

fix: enforce the public source artifact boundary#225
abrichr merged 3 commits into
mainfrom
fix/source-boundary-carve-20260721

Conversation

@abrichr

@abrichr abrichr commented Jul 21, 2026

Copy link
Copy Markdown
Member

What changed

  • Removes deployment-derived identity/reliability corpora, tuned evaluation material, and target-specific recipes from the public source and package boundary while retaining public runtime mechanisms, fake-patient demos, bounded aggregates, and reusable harnesses.
  • Adds a reviewed path-and-SHA-256 inventory for public data, evidence, static, model, and configuration payloads.
  • Validates current source plus embedded wheel/sdist inventories and rejects neutral renames, nested private paths, nested build/dist payloads, special files, unknown artifacts, tampering, and divergent wheel/sdist manifests.
  • Validates recovery builds against the target ref's embedded historical manifest under current policy instead of comparing old payloads with current-main hashes.
  • Corrects the paper's evidence contract and documents the already-public-history boundary.
  • Adds a corpus-free identity-rate scoring interface.

Why

The earlier denylist could be bypassed by neutral filenames and could break publish-existing-ref after legitimate artifact hashes changed. The positive inventory makes payload admission explicit while preserving recovery publication.

The inventory is intentionally payload-file coverage, not general DLP for arbitrary Python, Markdown, or TeX source.

Validation

  • 46 release-boundary, identity-rate, and claims tests passed.
  • Release source/version consistency, docs consistency, and claims gate passed.
  • Paper artifact constants passed; full-report and workshop PDFs built.
  • Exact rebased wheel and sdist built and passed embedded/cross-inventory validation.
  • Bounded non-e2e suite before the final recovery-only adjustment: 2,606 passed / 30 skipped.
  • Fast identity ladder: 22 passed; exact numeric-MRN OCR regression passed.
  • One legacy ONNX corpus harness remains pathologically slow locally and did not complete inside a standalone five-minute bound; the changed source-boundary code does not touch its runtime mechanism.

abrichr added 3 commits July 21, 2026 09:55
Keep the compiler, conservative runtime, generic validation harnesses, synthetic public demos, and bounded aggregate evidence open. Remove grown identity corpora, tuning sweeps, detailed target recipes, and raw reliability rows from the public checkout, and make both source-tree and distribution validation fail closed if they return.
@abrichr
abrichr force-pushed the fix/source-boundary-carve-20260721 branch from 25df9a0 to 62abc1f Compare July 21, 2026 14:02
@abrichr
abrichr merged commit 7472d19 into main Jul 21, 2026
17 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant