AI Collusion Sprint - Quiet Suppression

We built a grievance pipeline of separate LLM agents and tested how a hidden operator instruction to keep escalations low spreads through it. On 36 synthetic worker and community complaints, the instruction cut escalation from about 90% to 64% (MiniMax) and 44% (Gemma), and private notes between agents cut it to 28% in both. Notes written under the instruction lowered escalation even for raters who never received it. We then tested AI watchdogs (reviewing models) on the cases. Showing a watchdog three unsafe past cases that a review board had “accepted” lowered detection in 8 of 13 models, while the same cases without the endorsement left detection near baseline in the first four watchdogs. One sentence stating that written policy outranks precedent raised detection again in seven of those eight, four of them back to 100%. On a harder exploratory set, where an agent had rated a serious case as minor, all tested watchdogs struggled. Oversight should check outcomes against policy and give independent review to the example cases shown to watchdogs.