Recomputation

Recomputation checks a claimed computation by re-running it, or a random sample of it, from the same inputs on hardware the verifier trusts, and comparing the results with those reported ^{1 2}.

Re-running a large training run in full would be very costly, so in Shavit's framework the verifier re-runs selected segments, starting from a reported checkpoint and applying the reported data batches, and accepts if the result is close to the next reported checkpoint; this is the basis of proof-of-learning and training-transcript verification ¹. For inference, a trusted reference implementation recomputes what the model should have predicted at each generated token ², as in sampled inference recomputation; the same check can limit how much of a model's weights can be hidden in its responses (Bounding unexplained information in outputs) ³. Amodo Design distinguishes correctness, meaning that the workloads run match those declared, from completeness, meaning that every workload is reported, and notes that its recomputation schemes address only correctness ⁴. Recomputation also needs:

• Access to inputs and weights. This raises confidentiality problems, which Shavit addresses with a jointly trusted, air-gapped cluster ¹; one low-trust design uses air-gapped auditing environments with committed weights ⁵.
• Commitment before sampling. The prover must fix its records, for example by committing to a hash of sampled weights, before it learns which step will be audited ⁴.
• A way to handle numerical nondeterminism. Benign noise otherwise makes legitimate variation hard to tell from real problems ².