Early termination detection

An agent is partway through an open-ended task and something committed near the start may already have doomed it. You are designing the rule that decides when it stops pushing and starts over.

The task distribution is built to strip the easy signals. For every long run that is doomed there is a run of similar length and similar spend that succeeds, so elapsed steps and cost carry almost no information about the outcome. The agent's stated confidence is high in both, so self report carries almost none either. There is no external verifier to ask whether the trajectory is still viable, no terminal reward available mid-episode, and no per-task threshold: one rule with one parameter setting has to run across the whole distribution. Restarting is neither free nor clean. Some artifacts survive, a fixed restart cost is paid, and the context budget means the failed attempt cannot be carried back in full, so successive attempts are correlated in a way you do not control.

Deliver the revision rule: what it reads, what it computes, when it fires. Deliver the trade you are making between thrashing and sunk cost, as a bound against an oracle that abandons at the optimal moment if you can get one, or as an argument for why no such bound is available in this setting together with what you offer instead. Deliver the experiment that would show the rule fires on wrongness rather than on length, including the null it has to beat.

Ties break toward rules that remain non-trivial when the confidence signal is deleted from the inputs entirely, and toward rules whose behaviour under correlated restarts is analysed rather than assumed away.