Calibration in the Loop: Learning Early VLA Failure Detection with Condition-Wise Error Control under Deployment Shift

A failure monitor for a vision-language-action (VLA) policy is a learned score and a threshold that carries the false-alarm guarantee. Under deployment shift the guarantee must hold per operating condition, and the per-condition rule places the deployed threshold at the heaviest successful tail, where a score trained by cross-entropy alarms late, after a retry can no longer succeed. We frame monitor training under deployment shift as a constrained optimization problem: maximize early detection of failures subject to a false-alarm tolerance under every deployment condition, and we solve it with calibration in the loop (CIL), a training framework that computes the calibration rule that will set the deployed threshold inside the objective. The rule is an order statistic of reference successes per condition and the maximum over conditions, exact and differentiable almost everywhere, so the score is trained under the threshold it will be deployed with while dual ascent enforces one constraint per condition; the loss, the constraints and the rule can be exchanged. One trained score serves a marginal and a high-confidence deployment mode whose worst-condition bound holds with probability $1-\delta$. On held-out banks under the LIBERO-Plus catalog, CIL raises early detection within the tolerance from 39.3% to 59.2% for ${\pi }_{0.5}$ on LIBERO-Spatial and to 85.5% for OpenVLA-OFT when applied to a SAFE network, against detection baselines thresholded by the same rule; in a reset-and-retry loop the earlier alarms become completed tasks as the reset cost grows; and on a real xArm6 the monitor raises early detection from 34.9% to 44.4% within the tolerance and completes more tasks under recovery than the CoRe and TIDE monitors. Code and data recipes are available at https://anonymous.4open.science/r/CIL-4340/README.md.