Metamorphic Evidence as a Runtime Reliability Signal for Small Language Models

Small language models often run locally, where no larger model can check their answers, and the runtime exposes tokens and log-probabilities but not hidden states. MOME (Module of Metamorphic Evidence) is a label-free runtime gate. It edits each problem twice by deterministic rules, once keeping the correct answer and once changing a condition on which that answer depends. A predictor fitted offline on other backbones scores the answers. A two-stage policy serves high-scoring answers and verifies the rest by a vote over further samples or declines them. Mathematical word problems were used because every answer can be checked exactly. On five four-bit backbones and three benchmarks, the gate raised selective accuracy from 69.5% to 85.8% with two stage-2 samples and to 88.1% with four, at 4.34 and 5.70 generations per item. Correct yield stayed at the ungated level. The MOME score ranked answers with an area under the ROC curve (AUROC) of 0.863, against 0.734 for mean token log-probability. On four backbones of further families, it reached 0.829 against 0.629, and correct yield changed by −1.81 percentage points at stage-2 temperature 0.7 and by −0.19 at 0.3. The fitted components apply to mathematical problems, but the procedure can be reused in other domains.