Definition
Plain language
A chain of reviewers where the cheap, fast one looks first and only passes hard cases upward.
As stated in the literature
A staged monitor design in which a cheap classifier screens each action and escalates only ambiguous cases to slower reasoning or tool-using stages; because any stage may clear an action, early-stage false negatives are terminal.
Also called: escalation ladders
Why it matters: It keeps monitoring affordable at scale, but because any single stage can green-light an action, the weakest and cheapest reviewer effectively sets the floor on safety.
For example, a fast, cheap classifier glances at a routine "list the files in this folder" command and clears it instantly, while a request to send data to an unfamiliar address gets handed up to a slower model that can investigate.
Heard on the show
“Second, the monitor stages are an escalation ladder, not a gauntlet.”Episode 269 — How a Forged Transcript Got Model Weights Past a Safety Monitor