- Release README, case study, notes, policy, citation, license and contribution guide.
- Eval card, failure taxonomy and grading rubric.
- Task schema and ten task specifications.
- Agent-visible portions of the three runnable fixtures.
- Primary and supplemental run cards and summaries.
- Environment, model-selection caveat and machine-readable run index.
- Sanitized result JSON and reduced trace excerpts for both valid matrices.
validate_tasks.pyfor an offline schema check.MANIFEST.sha256for every exported file.
hidden_tests/and stored grader output.- Reference solutions.
- Full raw event streams and stderr files.
- Attempt snapshots, because they contain generated cache artifacts and duplicate the selected patch evidence in the trace excerpts.
- Invalid run payloads. The adjudication index keeps their status and reason.
- Inspect and OpenAI adapter payloads, which belong to the source project and need a separate dependency review.
The local exporter passing its safety scan does not publish anything. A maintainer must inspect the generated bundle, choose the license, create a clean repository and approve the disclosed-example policy before a remote push.