Evidence Projection: One Authorization Boundary, Many Jurisdictional Presentations
Annotatsiya
This paper introduces the evidence projection model for multi-jurisdictional AI governance. Across regimes, evidentiary expectations increasingly attach to how controls operated on specific system behavior rather than to policy attestation alone. Uncoordinated compliance stacks, with separate governance processes per jurisdiction, fail structurally because controls that operate without a common authorization discipline allow policy state, evidence, and enforcement consequences to diverge between operative enforcement and jurisdiction-specific documentation. Multiple authorization functions may govern the same proposed action where their authority, ordering, combination rule, conflict treatment, and execution effect are explicitly composed; the failure mode is undefined composition, not multiplicity. The paper proposes an architecture in which a single execution-time authorization boundary produces tamper-evident authorization artifacts, implemented as Proof-Carrying Decisions under the Five Tests Standard (5TS), that project into jurisdiction-specific evidentiary formats. Under this model, and within a declared projection set, jurisdictional fragmentation becomes an evidence-format problem rather than a governance-architecture problem. Evidence projection does not establish that any jurisdiction requires a pre-execution authorization artifact or that a projected evidence package satisfies a legal obligation. The paper defines evidentiary projection and the structural properties required for projectable authorization artifacts: reconstructability by an independent third party from the artifact and its bound materials under a declared replay mode (State-Replay or Protocol-Replay), verifiability, established input origin, completeness relative to the declared projection set and the evidence closure declared for the decision, and boundedness. Each jurisdiction-specific projection must be source-bound, non-augmenting, claim-traceable, mapping-versioned, independently regenerable, and scoped; projection is not indiscriminate replication, and selective presentation must remain traceably bound to the source artifact. The paper situates ALLOW / DENY / ABSTAIN verdict semantics within execution-time authorization and locates the completeness questions across the corpus instruments: the Authorization Artifact Test as the threshold, the Authorization Boundary Integrity Model separating Output, Input, and Replay Integrity, the Five Tests Standard as the normative test vocabulary (STOP, OWNERSHIP, REPLAY, ESCALATION, PROVENANCE), and the ABIM Evidence Requirements as the evidence standard. Provenance is treated as origin, not truth; it is necessary to Input Integrity but does not by itself establish admissibility. The paper further examines implications for enterprise governance architecture, infrastructure integration, audit consistency, and regulatory scalability across overlapping AI governance regimes, citing the EU AI Act, current US state automated-decision-making legislation, and Chinese algorithm-governance instruments as regime examples. It builds on prior FERZ work concerning execution-time authorization, tamper-evident authorization artifacts, fail-closed governance, and the distinction between observability, alignment, and authorization. Keywords: AI governance, AI compliance, runtime authorization, execution-time authorization, evidence projection, tamper-evident authorization artifacts, fail-closed governance, jurisdictional AI compliance, AI regulatory fragmentation, Five Tests Standard, Authorization Artifact Test, Authorization Boundary Integrity Model, deterministic AI governance, authorization artifact, independently reconstructable verdict. Version 2.0.0 (August 2026) retitles the paper and narrows its regulatory posture, replaces the parallel-boundaries claim with explicit composition conditions, adopts reconstruction under declared replay modes, scopes completeness to the declared projection set, adds projection invariants and selective-disclosure constraints, locates integrity questions across the Authorization Artifact Test, ABIM, 5TS, and the ABIM Evidence Requirements, incorporates the Override Asymmetry requirement that a replacement verdict materially consume the original decision materials, revises the main figure to show the authorization artifact and its bound materials as the evidentiary source, and adds official citations for the EU AI Act, Colorado Senate Bill 26-189, and the CAC algorithmic-governance instruments.
Hali tarjima qilinmagan