Burhan Engine 0.10.0 turns proof and evaluation into a stronger evidence gate.
Highlights:
- Public pytest proof runtime pinned by immutable GHCR digest.
- Safe sandbox verification for unified diffs from Aider, OpenHands, Copilot, and other tools.
- Conservative repairs for AttributeError, local import typos, literal KeyError, and TypeScript/JavaScript undefined identifiers.
- Separate diagnostic accuracy, exact patch success, diagnostic false positives, and patch false positives.
- Pinned reproducible BugsInPy experiment: 1/1 external fail-to-pass patch verified at V2 and 0/3 patch false positives.
Verification:
- 641 tests and 25 subtests passed; coverage 89%.
- Internal benchmark: diagnostic 100%, patch 4/4, diagnostic false positives 0%, patch false positives 0/9.
- Python 3.11/3.12, Docker, CodeQL, PyPI trusted publishing, and GHCR completed successfully.
License:
Version 0.10.0 is published under LicenseRef-Custom-Attribution with the complete terms in LICENSE. Earlier artifacts retain their originally declared terms; this release does not retroactively relicense them.