Language: English | 日本語

AI governance framework mapping

Status: Informational, time-bound (2026 Q2 reading) Date: 2026-06-12 Maintenance posture: This directory decays. The ADRs do not.

Purpose

The ADRs in this repository record judgments about how attribution should be distributed in autonomous AI agent systems. National and international AI governance frameworks (NIST AI RMF, ISO/IEC 42001, EU AI Act, Singapore's MGF for Agentic AI, OECD AI Principles) ship the governance framework layer — control catalogs, management-system requirements, risk tiers, practice catalogues, principle statements — that those judgments can be read against. This directory tracks, at the time of writing, how AAP's ten ADRs and four Business AI Quadrants relate to specific clauses of each framework.

The mapping is kept in a separate directory, not embedded in the ADRs themselves, to honour the repository's core thesis: implementation dissolves; judgment persists. Frameworks will be revised, superseded, re-numbered, or replaced; the ADRs will not. Editing this directory as frameworks evolve does not require touching the ADRs.

This is the governance framework layer counterpart to the industry mechanism layer mapping, which maps vendor products (Microsoft Agent 365, Agent Governance Toolkit, Purview, etc.) to ADRs. The two directories are siblings; they target different reader audiences (policy / compliance officer vs engineer / architect) and decay at different cadences.

Files in this directory

File	Framework	Framework version	Maintenance cadence
nist-ai-rmf.md (+ ja)	NIST AI Risk Management Framework	RMF 1.0 (2023-01) + Generative AI Profile NIST.AI.600-1 (2024-07)	yearly (GAI Profile expected to revise more often than core RMF)
iso-iec-42001.md (+ ja)	ISO/IEC 42001 AI Management System	ISO/IEC 42001:2023	multi-year (formal ISO revision cycle)
eu-ai-act.md (+ ja)	EU AI Act (Regulation (EU) 2024/1689), as amended by Regulation (EU) 2026/1744 (Digital Omnibus on AI)	in force 2024-08-01; Omnibus in force 2026-07-27; phased application through 2028	per delegated/implementing act, harmonised standard, or Commission guideline; else yearly (faster decay through the 2026–2028 phase-in)
singapore-mgf-agentic.md (+ ja)	Singapore Model AI Governance Framework for Agentic AI (IMDA)	v1.5 (2026-05-20, updated 2026-06-05); first released 2026-01-22	per MGF version (IMDA calls it a living document; v1.0→v1.5 took under four months — fastest decay in this directory); else yearly

The OECD AI Principles mapping is deferred to a later release (its decay cadence and granularity differ enough from the regulation/standard files that bundling it in would create churn).

Direction convention

ADR-centric primary direction. Each framework file lists ADRs 0001–0010 in order, and notes which clause(s) of that framework the ADR's judgment instantiates. AAP-side anchors stay stable across framework revisions.

Reverse index secondary direction. Each framework file ends with a reverse index (framework clause → applicable ADR) so that a reader coming from the framework side — a NIST control owner, an ISO 42001 auditor, an EU AI Act compliance officer — has a citation entry.

Framework version pinning convention

Each framework file's header pins the exact framework version / profile / publication date it was written against. Example:

Framework version: NIST AI Risk Management Framework 1.0 (NIST.AI.100-1, 2023-01) + Generative AI Profile (NIST.AI.600-1, 2024-07).

When a framework releases a revision, only the file targeting that framework is updated. The ADRs are not touched. Cross-version diffs in the framework do not propagate into the ADR bodies under any circumstance — see Revision procedure below.

Revision procedure

When a framework publishes a revision:

1. Open the framework file (e.g. nist-ai-rmf.md) and update its header Framework version: and Date: lines.
2. Walk each per-ADR mapping; revise wording where the new framework version moved, renamed, or retired the cited clause.
3. Update the reverse index.
4. Update the Japanese mirror to match.
5. Do not touch any file under docs/adr/. The ADRs' job is to record the judgment, not to track which framework cites what this quarter.
6. Commit with a message naming the framework + version (e.g. docs(policy-mapping): refresh against NIST GAI Profile 2026-Q4 revision).

If a framework introduces a clause AAP's existing ADRs have no judgment about, the ADR side is not extended retroactively — either a new ADR captures the judgment from first principles (with its own Lineage), or the mapping file notes the gap explicitly. Mapping coverage is not a target.

Non-attestation disclaimer (template — applies to every file in this directory)

This is AAP's reading. Each per-ADR mapping below records how AAP reads the relationship between one of its ADRs and one or more clauses of the named framework. It is not a compliance attestation: AAP does not certify that adopting an ADR satisfies a framework clause, nor that adopting a framework clause satisfies an ADR. It is not legal advice. The authoritative interpretation of any framework remains with the framework body (NIST, ISO/IEC, the European Commission, the OECD) and with qualified legal / compliance counsel for the deploying organization's jurisdiction. The mapping is offered as a citation surface for cross-reference, not as a substitute for the framework text itself.

Each framework file repeats this disclaimer in its own header.

Sibling: industry mechanism layer

../industry-mapping.md is the vendor product counterpart of this directory. It maps shipping vendor mechanisms (sub-millisecond policy gates, agent-identity primitives, sponsor systems, cross-vendor audit, registry sync) to ADRs. The two layers populate different parts of the AAP landscape:

• Governance framework layer (this directory) — what regulators, standards bodies, and management-system audits cite. Decays on framework revision cycles (yearly to multi-year).
• Industry mechanism layer — what vendor products ship. Decays on product release cycles (quarterly or faster).

A given ADR may have entries in one, both, or neither layer at any given time. The ADR itself remains the durable anchor.

Empirical corroboration

The per-framework files above read AAP's judgments against governance-framework text. A separate and weaker question is whether field evidence points independently at the same judgments. Through 2026, several enterprise surveys converged on a finding adjacent to AAP's thesis: the dominant cause of AI-agent production failure is the absence of accountability structure, not a deficit of model capability. They are listed here as external citation surface, in the spirit of the non-attestation disclaimer — they are not part of the AAP ADR lineage (the ADRs were extracted from implementation, not from these surveys), and survey methodology and numbers decay faster than the framework text the rest of this directory maps.

• Digital Applied, AI Agent Scaling Gap, March 2026. A 650-respondent enterprise survey: 78% run at least one agent pilot, only 14% reach production scale, and five organizational/operational gaps (legacy integration, output quality at volume, monitoring absence, unclear ownership, domain data) account for 89% of scaling failures. Organizations that stood up an operations function before any incident saw materially fewer production rollbacks. This corroborates the design→operation optimization-axis inversion behind ADR-0010 (phase separation) and the pre-named gap-bearer commitment behind ADR-0008 / ADR-0009: the accountable owner is named before the failure, not appointed after it.
• Grant Thornton, 2026 AI Impact Survey. Of nearly 1,000 senior leaders, 78% lack full confidence they could pass an independent AI-governance audit within 90 days, only 11% would prioritize risk and compliance as an AI success factor, and among the roughly three-quarters operating autonomous systems only about one in five has tested a response plan for AI failures. This corroborates ADR-0009 (triage before autonomy): an un-triaged deployment cannot answer the accountability question on demand.
• CSA, AI Agent Governance Framework Gap, 2026. Of 235 large-enterprise security leaders, 92% are concerned about AI-agent security yet most report significant governance gaps, 95% doubt they could detect or contain a compromised agent, only 38% monitor AI traffic (prompts, tool calls, outputs) end-to-end, and only 17% continuously monitor agent-to-agent interaction. This corroborates ADR-0006 (causal traceability): the infrastructure to reconstruct which agent did what, on whose instruction, under which credential is largely absent in the field.

These surveys are convergent, not foundational. AAP reached the same judgments inductively, from implementation friction; the surveys arrived at an adjacent conclusion independently and later. Treat the numbers as a 2026 snapshot that will decay; the ADRs they point at will not.

What this directory does not claim

• That AAP's ADRs cover every clause of any listed framework. The ADRs were extracted from implementation; the frameworks were drafted from policy. Coverage is partial in both directions.
• That a framework's clauses fully express the AAP ADR they map to. The judgment layer (when, why, at what cost) typically goes further than the framework's clause text.
• That the listed frameworks are the only relevant ones, or the most authoritative ones for any specific jurisdiction. Selection reflects current LLM-mediated citation surface, not a normative ranking.

Caveat

The frameworks listed here will be revised, renumbered, superseded, or replaced. The ADRs they map to will not. Update this directory as frameworks evolve; do not back-propagate framework changes into the ADRs.