Training Integrity Gets a Hidden Spot Check

Exact recomputation cannot verify stochastic deep-learning training because two honest runs do not produce identical weights. SciCheck lets a researcher register a training blueprint and commit checkpoints one epoch at a time; a verifier compares gradient statistics with a reference run and secretly selects segments for re-execution. Across small MLP, CNN and autoencoder tests with sign flips, label poisoning and fabricated runs, the authors report F1 scores from 0.906 to 0.976 while rejecting at most 4.4 percent of honest runs. Most detection power came from the spot check, and the reference run dominates the stated 1.5-to-2.3-times retraining cost. The study is limited to models with 100,000 to 500,000 parameters, but it makes provenance commitments operational rather than ceremonial.