ABIM Evidence Requirements: Evidence for Output, Input, and Replay Integrity Conclusions. A Supplement to The Enterprise AI Governance Buyer's Guide

This record contains ABIM Evidence Requirements, Version 3.5, a supplement to The Enterprise AI Governance Buyer's Guide (concept DOI 10.5281/zenodo.18002693). The supplement states what "demonstrated" means for the three properties of the Authorization Boundary Integrity Model (ABIM, DOI 10.5281/zenodo.20929115) when a procurement evaluator records, as the Buyer's Guide and its Procurement Fast Path instruct, whether Output, Input, and Replay Integrity are demonstrated. The problem it addresses. ABIM identifies conditions whose demonstrated failure defeats an integrity claim. It does not specify the affirmative evidence sufficient to support that claim. The corpus already carries most of the evidence machinery: the Enforcement Test Protocol in Observability Is Not Enforcement, the Authorization Artifact Test, the Five Tests Standard, and the Evidence Package and anti-laundering tests of the Buyer's Guide. What it did not carry was a property-level sufficiency standard, and for Input Integrity in particular, the affirmative evidence requirements. What the supplement adds. A single evidentiary threshold applied identically to all three properties: a property is demonstrated only when the required declaration is complete, direct evidence corroborates each material element of the claim, testable controls have been exercised under positive and negative cases, no confirmed failure witness exists, and no unresolved omission could materially affect the result. A supporting conclusion requires at least one negative case, selected or materially modified by the evaluator, for each control whose refusal behavior is material to it; vendor-selected cases alone do not meet the requirement, and the absence of a required evaluator-selected case leaves the property not established. External attestations, certifications, and policy documents standing alone are not direct evidence of what the boundary established. The vendor declaration on which every assessment rests: authorization object, action vocabulary, execution topology, decision state, evidence closure, replay mode, and temporal boundary. Scope follows covered effects, not vendor labels; any path reachable in the declared deployment that can produce a covered effect is part of the topology, an unlabeled reachable path leaves Output Integrity not established, and an authorization object narrower than the governed use leaves every property not established for that use. The affirmative evidence requirements for Input Integrity: identification of the evidence the verdict materially depended on, under a stated test of material use; admissibility conditions declared and traced to governing policy and authority, with not-applicable dispositions recorded; per-condition evidence across source identity, authority, integrity, freshness, permitted use, delegation and revocation, and temporal state; the evaluation record, with stated minimum contents, showing that the boundary tested each condition before emitting the verdict; and set-level admissibility, since conditions are relational rather than element-wise. Failure witnesses and not-established states for all three properties, kept separate: a confirmed counterexample defeats the claim; insufficient evidence leaves it not established and is never recorded as failure. Six independently recorded replay predicates per bound dependency: resolution, authentication, availability, executability, independence, and closure. A component that materially depends on live or undeclared state is not fixed by its own hash. Bridges from the Fast Path recording step to the instruments that carry the Output and Replay questions, and a coverage map showing what each Five Tests Standard test contributes to each ABIM property and what the test alone does not establish. An illustrative worked packet applies the procedure to a hypothetical payment-release system. It records a supporting Output Integrity conclusion, a not-established Input Integrity conclusion, and a defeated Replay Integrity claim within the same declared scope, demonstrating how the three property findings remain separate. What it does not do. The supplement is not an ABIM conformance profile. It issues no aggregate conformance determination, defines no accreditation requirements, authorizes no class of evaluator, and maps no evidence conclusion to the ALLOW, DENY, ABSTAIN verdict space. Conclusions are per property and descriptive: the evidence supports the claim, a failure witness defeats the claim, or the claim is not established. A supporting conclusion establishes that the evidence meets the threshold within a declared scope. It does not establish that the verdict was correct, that the evidence was true, or that the property holds outside that scope. Every recorded conclusion identifies the evaluator class. The Guide's weighted scorecard is unchanged. FERZ, Inc. makes no claim to satisfy these requirements. Doctrine boundaries carried from the corpus. Provenance establishes origin, not truth; it is one necessary component of Input Integrity and does not establish it. Input Binding is a substrate; hashing, signing, and immutable storage establish record integrity and answer neither the Input question nor, by themselves, the Replay question. Any human-intervention or override path capable of producing the governed effect is part of the execution topology, and authorized resolution remains inside the runtime authorization boundary. Machine-checkable Provenance conformance remains deferred under the Five Tests Standard; Input Integrity is assessed by evidence-backed human review against the threshold. The runtime authorization boundary is distinct from the RMF authorization boundary of NIST SP 800-37. Appendices. Appendix A is the boundary integrity worksheet that feeds the Fast Path Scoring Summary, with vendor and evaluator columns. Appendix B states the seven Input Integrity conditions and the failure witnesses and not-established states for all three properties. Appendix C is the 5TS coverage map. Appendix D is the worked packet. Relationship to the Buyer's Guide. This supplement is cited from Buyer's Guide v3.5 (Part II, boundary integrity section, and reference [10]) and from Procurement Fast Path v3.4 (Scoring Summary). It is deposited separately so that it carries its own citation and its own license. The Guide and the Fast Path carry their own terms. Audience. Procurement teams, risk officers, auditors, technical evaluators, and vendors preparing evidence for an Authorization Boundary Integrity assessment. Published by FERZ, Inc. under CC BY 4.0. The supplement is vendor-neutral and architecture-agnostic. It operationalizes the Authorization Boundary Integrity Model (DOI 10.5281/zenodo.20929115), the Enforcement Test Protocol in Observability Is Not Enforcement (DOI 10.5281/zenodo.18663864), the Authorization Artifact Test (DOI 10.5281/zenodo.20013582), and the Five Tests Standard (concept DOI 10.5281/zenodo.21040295), and applies The Override Asymmetry (DOI 10.5281/zenodo.19772248) and The Closed-World Bargain (DOI 10.5281/zenodo.21643658) at their stated limits.

Hali tarjima qilinmagan