Commit Before Binding: Typed Effect Programs for Held-Out Tool Families

Tool interfaces certify call shape, not state-change intent: a schema-valid agent can still mutate the wrong resource, cross an authorization boundary, or commit an irreversible transition. RAVEL transfers the missing invariant as a provider-independent typed effect program over resource identity, validation, permission, mutation scope, and recovery; provider operations and arguments become available only after the program composes, while uncertain critical bindings trigger read-only probes or pre-mutation abstention. We evaluate this commitment-first architecture on 612 resource-mutating tasks in six leave-one-family-out folds, exposing target documentation and schemas under a common 15-call budget while withholding target-family trajectories and supervision. Against same-backbone Adapter Transfer, RAVEL raises fold-mean success from 40.3% to 48.3% (paired gain 8.0 points, 95% CI [6.1, 9.8]) and lowers task-weighted catastrophic error from 10.1% to 4.7% (RAVEL-minus-Adapter CI [-6.9, -3.8]). A latent-plan control matched for source data, annotation and verification effort, capacity, corruption training, hard negatives, probing, and abstention attributes a further 3.2-point success gain and 2.1-point catastrophic-error reduction to the joint typed structure. All six familywise catastrophic-error intervals favor RAVEL; under misleading descriptions it retains 77% of clean completion versus 64% for Adapter Transfer, and an independently assembled 184-task AppWorld-derived suite reproduces the ordering (44.2% versus 40.7% for the matched control and 35.9% for Adapter Transfer). Familywise replay identifies scheduler, dependency, trigger, and retry state as the residual representation bottleneck. Making commitments executable before an unfamiliar call can alter state therefore improves completion and mutation integrity while revealing where resource-centric effects require temporal structure.