Generating plain-English description…
Plain-English summary of the active rules — what agents can do, what gets blocked, and what needs a human.
This policy applies only to the `research_crew` agent in the `dev` environment. It may use only the tools `arxiv_search`, `web_search`, and `delete_research_data`, and it may use only the model `gpt-4o-mini`; however, `delete_research_data` requires approval before it can run. Each request is limited to 4,000 tokens, with warnings at 80% usage, and overage is tightly controlled with a hard deny at 2× the limit, a maximum of 3 overages in a 24-hour UTC window, and a repeat limit of 3. Safety guardrails require PII to be redacted, and any prompt injection attempt or data exfiltration attempt must be denied.