This preprint presents a verifiable contract framework for modular pipelines in AI and general composite systems under two strict constraints: only observable evidence is admissible, and no privileged external meta-evaluator is assumed. The model is domain-agnostic and applies to AI pipelines, software supply chains, cyber-physical control stacks, and scientific workflow graphs. The paper introduces three machine-checkable certificate profiles—COMPOSITIONCORE, TRUEBRIDGE, and FULL—that progressively add (1) compositional core guarantees, (2) audit-calibrated proxy-to-true risk accounting, and (3) drift hardening with typed evidence semantics. A deterministic, fail-closed verifier recomputes declared bounds from canonicalized payloads, dependency budgets, and signed campaign evidence, rejecting on any mismatch. Formally, the work provides nonlinear local module contracts, high-probability global composition bounds (quality, compute-path, compute-total, latency, proxy-gap, boundary interaction, and residual risk), explicit claim-scope semantics, and non-vacuous confidence budgeting. Practically, it includes schema-exact JSON constraints, executable verifier pseudocode, assumption-to-field traceability, exact binomial falsification design with Type I/II targets, numeric precision tiers, conversion profiles, resource caps, and Byzantine-aware signature/keyset validity policies. Category-theoretic operators are used for boundary-aware composition and recertification-efficient module evolution, linking formal guarantees directly to implementation obligations in adversarial multi-component environments.
K Takahashi (Mon,) studied this question.