Skip to content

ISO/IEC Standards Rule Engine — Post-Seed Plan

Status: Deferred from v0.1. Planned post-seed. Scope owner: Policy + Engine. Source of truth: PRD §16 stack (rule engine), scope-90-day.md:36.

ISO/IEC 42001 (AI management system) and ISO/IEC 23894 (AI risk management) are management-system standards rather than line-by-line technical controls. The product's role is to surface deterministic evidence for the clauses that are technically observable, and to emit structured HITL prompts for the rest. This is a rule pack and an evidence-mapping table — not a new engine.

Why it is deferred

ISO/IEC standards are paid documents with audit ecosystems that move on year-scale cycles. They are not the wedge for design-partner activation. Shipping them before the EU pack is battle-tested in CI would invest engineering against the wrong pressure curve. Post-seed, ISO becomes a procurement-unlock for enterprise deals.

Architectural preconditions v0.1 must preserve

The same set as for the NIST pack applies. The additional ISO-specific invariants:

Precondition Owner phase Why it matters
Evidence ledger supports linking a single event to multiple control IDs across frameworks Phase 7–8 One control in EU AI Act may map to several ISO 42001 clauses; evidence must not duplicate
Annex IV dossier metadata envelope is extensible to additional document types (ISO statements of applicability, risk registers) Phase 12 The dossier generator becomes a multi-document pipeline
Override/HITL rationale capture is structured (controlled vocabulary) where possible, not free text Phase 11 ISO audits require reproducible evidence of management decisions
Policy bundle schema supports "non-technical" controls that resolve only via signed HITL attestations Phase 10–11 Many ISO clauses are management commitments, not measurable system facts

Scope when it ships

In scope:

  • ISO/IEC 42001 control mapping: clauses 4–10 mapped to a mix of deterministic checks and structured HITL attestations.
  • ISO/IEC 23894 risk-management process mapping: aligned to the existing Risk Engine outputs (no new risk taxonomy).
  • Statement of Applicability (SoA) generator as an output adapter against the evidence ledger.
  • Risk register exporter aligned to ISO 23894 vocabulary.
  • Cross-mapping table EU AI Act ↔ ISO/IEC 42001 published as data, surfaced in the Annex IV dossier when both targets are selected.

Out of scope (intentional):

  • Certification audit support. The product produces evidence; certification bodies remain external. No "auto-certify" claims, ever.
  • ISO/IEC 27001 / 27701 — these are organizational security and privacy standards; separate scope.
  • Embedding the standard text. License terms forbid it; the product references clauses by ID and the customer holds their own copy.

Milestones (indicative, post-seed)

  • I0 — Mapping table published as data; SoA adapter scaffolded.
  • I1 — Clauses 4–7 (Context, Leadership, Planning, Support) deterministic where observable; HITL prompts where not.
  • I2 — Clauses 8–10 (Operation, Performance evaluation, Improvement) covered; first enterprise customer runs ISO + EU dual-target.

Exit gate for "GA": SoA adapter passes a third-party reviewer round; at least one enterprise customer uses the ISO output in an actual external audit.

Open questions

  • Whether to license the standard text via an audit-partner relationship or stay reference-only.
  • Whether ISO 42005 (AI impact assessment) is in the same bundle or its own pack.
  • How ISO certification timelines interact with our release cadence — likely we publish quarterly "audit snapshot" tags.