Less guesswork. More working software.
A good README should be a path from a fresh machine to a useful result.
Why agent-ready matters
Finding a promising repo is easy. Finding out whether an agent can actually use it takes time. Coldpass records that first attempt so you can see what worked, what got stuck and what remains unknown before trying it yourself.
Cheap agents expose friction
We start with Haiku, with Luna as another observation where available. If a small model can follow the instructions, the project is easier to pick up. If it struggles, the transcript can point to a missing prerequisite, unclear instruction, test problem or product bug. The checker must confirm the result.
Useful to users and maintainers
Users get short answers with evidence. Maintainers get a reproduction and, when we can verify one, a patch. A test failure is investigated before it becomes a claim about the project.
Evidence earns trust
Keep the task, revision, test version and limitations attached to every answer. Make fixes reproducible. Let maintainers correct the record. Nobody pays for a better verdict or rank.