SpliceGuard: Composition-Time Contracts Recover Strict Inference Budgets
Abstract
Terminal validation can detect a structural defect but cannot refund calls spent extending invalid intermediate state. SpliceGuard turns certificates into composition-time control: interface refinement governs admission, one bounded repair can restore a failed boundary, and dependency-aware replay recomputes only the affected suffix. Freeze-then-replay holds candidate membership, executor, cache policy, accounting, and replay order fixed to identify budget recovery. Under GPT-4o-mini (2024-07-18) and a 10k-call limit, replaying 38,482 candidates raises candidate-level final-checker accuracy from 72.3% to 82.7% with contract admission and to 85.6% with repair. A second 27,346-candidate replay compares terminal validation, eager checks, full-rerun schema gating, contract admission, and repair under the same instance-level success denominator: gate-only reaches 75.8% and gaterepair reaches 78.1%, versus 64.1%, 69.2%, and 71.3% for no filtering, eager validation, and a full-rerun schema gate. The comparison isolates refinement authority and replay locality beyond earlier detection. A structural refinement proposition and a fixed-pool throughput condition specify when early checking can convert avoided late work into usable budget. Composition-time contracts thereby make strict-budget workflow composition an auditable control mechanism.
est. 32% chance this paper gets accepted at ICLR 2027.
What do you think this paper will get?
All positions stay anonymous.