When Grasp Metrics Disagree: Contact, Load, and Retention in Simulation

A simulated object can touch its environment without receiving load, and can receive load without depending on it for a specified continuation. Grasp evaluations that conflate these properties can disagree without either implementing its stated rule incorrectly. We separate literal no-contact, realized no-external-load, and intervention-conditioned retention, and test their consequences using frozen protocols in MuJoCo. Across 240 free-object controls, no-contact rejects 56 of 77 no-load reference positives; eight representation-only pairs change its verdict despite identical full-state trajectories. However, an integrated load-and-geometry monitor provides no observed gain over a full-wrench baseline on either registered primary comparison. It rescues none of 85 historical excluded scenes. On 120 fresh Allegro scenes, both monitors make identical selections or abstentions: 13 returns, 10 successful independent 15-second continuations, and 107 abstentions. Native-gate selection returns 40 candidates with 15 successful continuations, exposing a failure-versus-coverage tradeoff rather than uniform improvement. Fetch and Shadow adapters further reveal task and uncertainty boundaries. The result is a measurement study with explicit negative findings, not a superior grasp evaluator: evaluation targets, intervention horizons, and abstention costs must be stated before acceptance rates can support a claim.