The three executable v0 fixtures are disclosed examples. A public bundle may include:
- task prompts and specifications;
- agent-visible source and public tests;
- redacted model trace excerpts;
- aggregate and task-level outcomes;
- descriptions of the hidden behaviors exercised in the case study.
The bundle excludes:
- hidden-test source files;
- raw grader output and assertion traces;
- reference solutions;
- raw OpenCode event streams;
- local paths, session IDs and cache artifacts.
The case study describes enough behavior to contaminate the current hidden checks for future model evaluation. Any result collected after publication must label these tasks as disclosed examples. It cannot present them as an unseen holdout.
A new private set should use different tasks and invariants. The maintainer should keep its prompts, grader inputs and hashes outside the public bundle, record access and freeze the evaluator before collecting runs. Publication can release aggregate results and selected failure excerpts while retaining the new test source.
The exporter uses an allow-list. It creates short trace excerpts from text, edit, write, bash and tool-error events. It removes internal IDs and rewrites private paths. A scan fails the export if a user path, username or raw OpenCode ID key remains.