Threshold Choice, Not Sample Size, Bounds Trustless Verification of Nondeterministic Compound AI Workflows
A protocol for verifying nondeterministic compound AI workflows records hashes of each stage’s inputs, outputs and context and settles challenges via re-executions without a quorum. In a synthetic HotpotQA pipeline, a threshold derived per execution detected 19 of 29 fabrication trials versus 9 for the best fixed threshold matched to zero honest rejections, with no honest commitment rejected at k=5; the rule is measured, not deployed.