The Enterprise AI Governance Buyer's Guide

This record contains Version 3.5 of The Enterprise AI Governance Buyer's Guide, together with the companion Procurement Fast Path (Version 3.4). The two documents are designed to be used jointly to support rigorous, evidence-based evaluation of AI governance claims in regulated and high-stakes enterprise environments. Version 3.5 is accompanied by a separately deposited supplement, ABIM Evidence Requirements (DOI [supplement DOI, insert on deposit]), which states the evidentiary threshold behind the Guide's three boundary-integrity questions. The Buyer's Guide presents a vendor-neutral evaluation framework that distinguishes three governance problems: Visibility: logging, monitoring, and observability. Alignment: model- and application-layer safety techniques such as RLHF, guardrails, and content filtering. Authorization: pre-execution governance capable of producing replayable, independently verifiable evidence that a specific action was permitted. It formalizes the distinction between probabilistic governance (likely compliant) and deterministic governance (independently verifiable for any specific decision), with emphasis on fail-closed enforcement, non-delegable authorization, and post-incident verifiability. The verdict space is exactly three: ALLOW, DENY, and ABSTAIN. ABSTAIN blocks execution pending authorized human override. An authorized human resolution is a governed authorization event: execution may proceed only if the runtime authorization boundary materially consumes the authority-bound human input and emits the resulting action-bound ALLOW authorization artifact. Determinism is conditional on the declared governed state and the correct operation of the boundary; it establishes reproducibility of the verdict over that state, not that the state is complete, admissible, or a complete representation of the external world. The Five Tests Standard (5TS). The guide evaluates governance against the published, vendor-neutral Five Tests Standard v1.2.0 and its five canonical tests: Stop: execution can be halted before a side effect occurs, with effect-token issuance gated by approval and no execution pathway bypassing the boundary. Ownership: an identified authority signs the governing policy prior to the execution window, with attributable responsibility chains. Replay: the decision can be reproduced at the boundary by state or by protocol, including third-party offline reconstruction from the Evidence Package without access to the governed system. Escalation: mandatory custody transfer routes on denial or threshold crossings, with explicit human-in-the-loop routing. ABSTAIN blocks execution pending authorized human override; escalation is the consequence of ABSTAIN, not a verdict. Provenance: the origin of the inputs grounding the verdict is established. Provenance is origin, not truth: it does not by itself establish that input content is true, complete, or current. 5TS v1.2.0 defines five normative tests. The current machine-checkable conformance bundle (v1.0.2) covers Stop, Ownership, Replay, and Escalation; Provenance is evaluated as a normative requirement, and Provenance conformance and full five-test conformance are not yet assertable pending publication of the Provenance bundle and test vectors. 5TS Provenance addresses one necessary component of ABIM Input Integrity but does not establish Input Integrity by itself; Input Integrity additionally requires the boundary to evaluate and enforce the applicable admissibility conditions. Authorization Boundary Integrity Model. The guide incorporates the Authorization Boundary Integrity Model, Version 1.1 (concept DOI 10.5281/zenodo.20929115; version DOI 10.5281/zenodo.22104753), which treats output integrity, input integrity, and replay integrity as independent properties of a single authorization boundary. Holding one implies nothing about the others, and Input Binding, the recording and binding of the inputs a decision used, is the substrate beneath the three rather than a fourth property. Input Integrity is stated as admissible origin: no verdict rests on evidence whose admissible origin has not been established. The model is used to locate where a vendor's boundary holds and where it does not before mechanisms are compared. The per-property evidence procedure, including the evidentiary threshold and the affirmative evidence requirements for Input Integrity, is supplied by the ABIM Evidence Requirements supplement. The guide supplies concrete due-diligence questions, a five-point diligence test with an Input and Artifact Integrity prong, five anti-laundering tests, failure-mode analysis, and an illustrative tamper-evident authorization artifact, implemented as a Proof-Carrying Decision (PCD) under 5TS, mapped to common regulatory requirements such as audit trails, electronic records, signature controls, and retention obligations. A regulatory mapping identifies where pre-execution authorization may be relevant across the major regimes (EU AI Act Article 14, GDPR Article 22, HIPAA 164.312(a), DFARS 252.204-7012, and NIST AI RMF GOVERN). The mapping identifies potential architectural evidence relevant where applicable law, contract, or organizational policy requires authorization before action; it does not assert that these provisions create a uniform requirement for an independently reconstructable, action-level authorization verdict. An EU AI Act Evidence Support discussion describes how deterministic governance infrastructure can support control and documentation obligations, with applicability dependent on system role, risk classification, deployment context, and legal interpretation. Corpus Doctrine. Observability explains what happened. Enforcement determines what is allowed to happen. Enforcement is realized through a non-bypassable authorization boundary that emits a tamper-evident authorization artifact prior to execution. What is new in Version 3.5. The Appendix B definition of Input Integrity corrected to admissible origin, aligning the Guide with ABIM v1.1; two companion origin-only formulations corrected to admissibility. The override-artifact sentence removed from the ABSTAIN definition pending settlement of separate override-record requirements, and the human-resolution invariant stated in its place: an authorized human resolution is a governed authorization event in which the boundary materially consumes the authority-bound human input and emits the action-bound verdict. The liability-allocation entry for human override restated as Authorized Human Resolution. The regulatory mapping calibrated to potential architectural relevance, distinguishing what each provision requires from the architectural evidence that may support it, and distinguishing access authorization from action-level authorization. Determinism bounded to the declared governed state and the correct operation of the runtime authorization boundary. The relationship between 5TS Provenance and ABIM Input Integrity stated directly: Provenance is necessary for Input Integrity and not sufficient for it. The ABIM reference updated to Version 1.1, and the ABIM Evidence Requirements supplement linked from the boundary-integrity section and the References. The companion Procurement Fast Path aligned to Version 3.4; its Scoring Summary now points to the supplement for what "demonstrated" means for each integrity property. Both PDFs include a navigable bookmark outline for section-level access. Version lineage. 3.4 migrated the Guide from the four-test predecessor to the Five Tests Standard v1.2.0, adding Provenance as the fifth canonical test, adopted the Authorization Boundary Integrity Model as a three-property diligence lens, extended the Five-Point Diligence Test with an Input and Artifact Integrity prong, stated ABSTAIN semantics in full, and aligned conformance-claims discipline with the standard. 3.3 aligned the four tests with the canonical published predecessor standard and adopted the artifact gap designation from FERZ Technical Advisory TA-2026-01 (DOI 10.5281/zenodo.20646404): the evidentiary condition in which records of what occurred exist, but no authorization artifact preceding execution can be produced. 3.2 operationalized the Authorization Artifact Test for procurement, advanced the corpus Doctrine to v1.1, and added the regulatory mapping. 3.1 added a formal Doctrine (v1.0) statement locking the enforcement boundary across the corpus: observability governs accounts of action; authorization governs permission to act; and authorization governance requires a non-bypassable runtime gate that fails closed when governance conditions fail or required evidence is missing. 3.0 introduced semantic completeness: governed state is incomplete unless each decision is verifiably bound to the immutable semantic definitions in effect at decision time. This prevents definition laundering, where semantic drift retroactively alters the compliance status of past decisions and breaks audit replay without visible failure. Companion: Procurement Fast Path (Version 3.4). The Fast Path distills the framework into a three-page operational checklist for real-world procurements. It establishes a baseline qualification gate requiring third-party offline replay of a historical governance decision from an exported Evidence Package, applies the canonical Five Tests, and adds anti-laundering tests (exportability, offline replay, state completeness, fail-closed behavior, and mutation or drift resistance) to disqualify marketing-only or trust-based governance claims before full scoring. Its Scoring Summary records separately whether Output, Input, and Replay Integrity are demonstrated and refers the evaluator to the ABIM Evidence Requirements supplement for the evidentiary threshold and the per-property criteria. Version 3.4 is derived from Buyer's Guide v3.5. Audience. Procurement teams, risk officers, General Counsel, auditors, technical evaluators, reg