Which Value Head Missed? Outcome-Resolved Proper-Score Excess for World-Model Control

A world-model forecast can be uncertain without a consequential outcome, while a confident reward, cost, or termination forecast can resolve with unexpectedly large proper loss. The environment response therefore exposes a learning signal unavailable at forecast time: which value head realized loss above that head's self-expected score. We define outcome-resolved value score (ORVS) as the positive, nominally standardized proper-loss residual and introduce Outcome-Resolved Score Control (ORSC), which propagates immutable collection-time labels through a terminal-masked continuation critic into guarded actor updates. Under a common late-update path, ORVS raises shifted return over uncertainty by $43.9[30.5,57.3]$ points with similar nominal return; a separate ranking analysis finds that its top 10% contains $0.61\pm 0.04$ of subsequent harmful continuations, versus $0.49\pm 0.03$ for predicted cost and $0.42\pm 0.04$ for uncertainty. The control effect recurs with a TD-MPC-style agent ($36.6[23.4,49.8]$ shifted-return points) and unseen visual distractions ($0.744\pm 0.022$ retention), while SafetyPoint and SafetyCar cost rates reach 2.43 and 3.07 per 1,000 steps.