This is a useful early evaluation of monitorability, but its preliminary setup does not establish how reliably monitors would detect consequential real-world sabotage.
Limits of the evidence: The source describes early-stage results from a small, limited-diversity task set and notes a lack of agent and monitor elicitation; results are not independently corroborated in the supplied text.
What to watch
Taken from the first report; the whole story has not been re-assessed yet.
Early work on monitorability evaluations
METR BlogJan 22, 2026First report
No revisions yet.
No maturity ladder applies to this desk.
Sens.ai aggregates and assesses published reporting. The assessment above is machine generated; the original sources are authoritative.