FIRE: harness-injected runtime policies at pre-failure states lift Terminal-Bench pass^2 from 64.4% to 73.6% without touching the model

arXiv 2609.26048·high signalXBlueskyLinkedInCopy link

FIRE mines states that preceded observed agent failures and has the harness inject targeted instructions or action denials when those states recur. GPT-5.6 Sol's best-of-two barely moved (+1.2 points) while repeated success rose 9.2 points, so the policies turn solutions the agent can already reach into ones it delivers every time. In a randomized five-arm test, real policies scored 61% against 36-43% for a timing-matched sham and generic 'verify' or 'reconsider' nudges. Generic self-check prompts do almost nothing. State-specific rules taken from your own failure logs do.