Apollo published a monitoring-evaluation roadmap on November 19, 2026. It proposes research methods, not a completed audit or deployment safety result.
Six claims require different evidence
Its six claims concern coverage, detection, trustworthiness, disclosure, previously missed incidents and future adequacy. Coverage reconciles complete, tamper-resistant inference logs with monitoring records. Coverage alone does not establish an adequate monitor.
Proposed collusion and under-reporting tests cannot conclusively establish trust through negative findings. Containment is outside the post's main scope.
Detection includes the route to human review
Asynchronous detection includes human review. Apollo distinguishes flagging delay from review delay, making thresholds and available reviewer capacity relevant.
Engineering interpretation: counting alerts alone cannot show whether every run was observed or whether anyone examined an alert in time. A useful evidence record should make these separate questions answerable without relying on a single reassuring dashboard total.
A proposed audit with deliberately missing records
Original exercise, not executed: create 30 numbered synthetic agent runs. Leave five without monitoring records and delay human acknowledgement for two flagged runs. Declare the expected flag for each fixture before calculating any result.
Ask the learner to report coverage, detected fixtures and review delay separately. Then change only the acknowledgement times. Coverage should remain unchanged while the review result changes. This demonstrates why the measurements describe different properties.
The toy exercise would not test a monitor's resistance to strategic deception. It would make gaps in the evidence path visible before anyone used the coverage figure to support a broader assurance claim.
Sources accessed November 21, 2026. Views are my own and do not represent any employer, past or present.
Sources and evidence
Evaluate recorded coverage, detection quality and timely human review separately; a roadmap is a method to test, not a safety result.