One Retrieval Slice, Two Winners: Trajectory-Aware Groundedness Repair with GroundDial

The same 512-prompt Natural Questions (NQ) retrieval slice selects LoRA Steering under post hoc best-point Hit (0.853 versus 0.849) and GroundDial under the declared whole-trajectory utility (0.824 versus 0.742); this disagreement recurs in 78.2% of paired trajectory resamples. This reversal exposes the central evaluation problem: a fixed evidence pool still leaves the evidence object, operating-point rule, and latency envelope to determine which repair is preferred. A claim contract formalizes these choices, and GroundDial supplies a post-retrieval controller that combines trajectory-regularized low-rank hidden-state updates, an output-factor norm cap, and selective abstention. Under the primary NQ latency contract, GroundDial reaches 0.847 Hit and 0.062 Side, versus 0.758/0.112 for Cross-encoder Bo4; full-slice blinded experts measure support gains of 10.5 points on NQ and 8.8 on MS MARCO, with preservation rates of 0.887 and 0.859. Across the same 11 native-path positions, GroundDial obtains 0.912 monotonicity and 0.824 AUC, exceeding TruthX-style steering at 0.861/0.787 and LoRA Steering at 0.803/0.742; trajectory regularization accounts for the largest monotonicity ablation loss, while the norm cap and gate control Side. The ordering persists across three backbones and NQ contexts through 16K, while matched-latency advantages extend to MS MARCO, ASQA, and Contriever retrieval. Claim contracts turn fixed-retrieval evaluation from an endpoint leaderboard into a decision-specific test of selected quality, controllability, and workflow cost.