A small draft model scoring an agent's own trajectory predicts failure before execution.

This inverts speculative decoding. A small open-weight draft model scores a black-box agent's already-generated trajectory in one forward pass, needing no logits, weights, activations or repeated sampling (arXiv 2609.05274). Phase-aware features separating reasoning spans from action spans get calibrated against a verifiable objective into a failure-likelihood score. Wired into a pre-execution veto gate on Qwen3-Coder-480B and Claude 3.5 Sonnet it cut execution error rate 6-8 points and token cost 14-19%, and transferred to out-of-distribution benchmarks without retraining.