A fail-closed control layer and evals for autonomous agent fleets

Loss-of-control scenarios presuppose that AI systems act past the mechanisms meant to constrain them — yet today nobody can measure whether those mechanisms work. Every agent framework ships its own guardrails; documented incidents (a deleted production database, five-figure runaway spend) show they fail in ordinary use, and there is no reproducible way to ask: does this control layer actually stop an unsafe action? This project turns that question into measurement: a public eval spec with 20–40 reproducible scenarios across 5–6 risk categories, plus an automated red-team explorer that searches for boundary-crossing agent behavior in disposable sandboxes — all run against a real, operating control plane and published with the bypasses and failures included. The aim is to do for agent control what jailbreak evals did for model-level safety: make "we have guardrails" a testable claim, with block rates, false-positive rates, and a failure taxonomy, before autonomy scales. This does not solve alignment; it makes control of agent actions measurable — which any later containment strategy will need.