Skip to content

iso42001-soa.md

Latest commit

 

History

History
119 lines (98 loc) · 7.83 KB

iso42001-soa.md

File metadata and controls

119 lines (98 loc) · 7.83 KB
Raw

ISO/IEC 42001 — Statement of Applicability (Sprout)

Standard: ISO/IEC 42001:2023 (AI Management System), the 42 Annex A controls (groups A.2–A.10). This SoA records, per control, whether it applies and how Sprout satisfies it — or a justified N/A. Per STANDARDS/RESPONSIBLE-TECH-FRAMEWORK.md §Governance, this is a REVIEW-GATE artifact, regenerated on release and post-architecture-change.

  • Author / accountable owner: Chelsea Kelly-Reif (Cl. 6.2)
  • Last regenerated: 2026-06-22 · Review cadence: annual + on any architecture change
  • Certification status: not certified. Sprout is an independent open-source reference implementation; this SoA demonstrates the management-system spine as a public artifact, not a certification claim.
  • Scope: the Sprout assistant (offline-default + opt-in cloud seam) and its eval harness. Family Greenhouse personalization is deferred and excluded.

Numeric thresholds (coverage, axe impact, faithfulness floors, SHA-pin requirements) are owned by the sibling standards and linked, not restated here.

Annex A controls

Status key: A = applicable & implemented · N/A = not applicable (reason given).

A.2 — Policies for AI

Ctrl Control Status How / artifact
A.2.2 AI policy A The four hard rules (cite-or-refuse, never-certify-safe, dated corpus, offline-default) are the policy; in CLAUDE.md + README, enforced in code
A.2.3 Alignment with org policies A Apache-2.0; NOTICE independence statement; portfolio STANDARDS referenced repo-wide
A.2.4 Review of the AI policy A Quarterly, with the risk-register review; tied to framework recheck cadence

A.3 — Internal organization

Ctrl Control Status How / artifact
A.3.2 AI roles & responsibilities A Single accountable owner named (Cl. 6.2); CODEOWNERS gates the load-bearing guards
A.3.3 Reporting of concerns A SECURITY.md, issue templates, CODE_OF_CONDUCT

A.4 — Resources for AI systems

Ctrl Control Status How / artifact
A.4.2 Resource documentation A Architecture + repo layout in docs/ARCHITECTURE.md, README
A.4.3 Data resources A corpus/manifest.yaml (dated, licensed, synthetic CC0); datasheet for datasets (transparency audit)
A.4.4 Tooling resources A Pinned deps; SHA-pinned actions; SBOM on release (security standard)
A.4.5 System & computing resources A Offline default needs no GPU/cloud; cloud seam region/limits in config
A.4.6 Human resources / competence A Single maintainer; model card states reviewer competence assumptions

A.5 — Assessing impacts of AI systems

Ctrl Control Status How / artifact
A.5.2 AI system impact assessment process A Process defined in the framework; impact assessment scoped — see note below
A.5.3 Documentation of impact assessments A Risk register + EU AI Act classification committed; full AI System Impact Assessment deferred (Cl. 6.1.4 trigger not met: no personal data processed in default mode, no consequential decisions)
A.5.4 Assessing impacts on individuals/groups A Ethics + privacy audit narrative; worst-case = false "safe" → addressed by never-certify-safe
A.5.5 Assessing societal impacts A Low-stakes domain by design; documented in risk register R6/R7

A.5.3 note: the dedicated ai-impact-assessment.md is N/A for now — its Cl. 6.1.4 trigger (personal data, consequential decisions, or external-user exposure with risk) is not met by the offline default. It becomes applicable when the Family Greenhouse household-data path ships.

A.6 — AI system life cycle

Ctrl Control Status How / artifact
A.6.2.2 Objectives for responsible development A The four hard rules + the eval scoreboard are the objectives
A.6.2.3 Design & development processes A src/ module boundaries; ADRs for any change to guards/thresholds/loader
A.6.2.4 Verification & validation A make verify = lint·type·test≥90%·security·a11y·eval; eval harness is the V&V instrument
A.6.2.5 Deployment A pipx/container/one-command serverless; reproducible, content-hashed builds
A.6.2.6 Operation & monitoring A Tier C structured logs (Tier A for cloud API); health endpoint; debug retrieval trace (AnswerTrace)
A.6.2.7 Technical documentation A docs/ set: ARCHITECTURE, THREAT-MODEL, ACCESSIBILITY, model card, ADRs
A.6.2.8 Recording of event logs A Per-answer trace; eval runs content-hashed and byte-reproducible

A.7 — Data for AI systems

Ctrl Control Status How / artifact
A.7.2 Data for development/enhancement A Synthetic CC0 corpus; no scraped third-party data in default build
A.7.3 Acquisition of data A Manifest records source/url/license/fetch-date per doc
A.7.4 Quality of data A Cleaned, chunked, content-hashed; dated; tamper-evident index
A.7.5 Data provenance A Provenance tag (corpus) on every rendered sentence; citation carries doc_id + fetch_date
A.7.6 Data preparation A ingest.py fetch→clean→chunk→embed→index; deterministic

A.8 — Information for interested parties

Ctrl Control Status How / artifact
A.8.2 System documentation for users A README quickstart, /help, example questions, disclosure on every answer
A.8.3 External reporting A SECURITY.md disclosure path; public issue tracker
A.8.4 Communication of incidents A CHANGELOG + dated audit regeneration; red-team reports committed
A.8.5 Information for interested parties (limits) A Model card states intended/out-of-scope use and what Sprout cannot do

A.9 — Use of AI systems

Ctrl Control Status How / artifact
A.9.2 Processes for responsible use A Disclosure + "not veterinary advice"; honest refusal as first-class output
A.9.3 Objectives for responsible use A Cite-or-refuse and never-certify-safe enforced at output, not left to the user
A.9.4 Intended use A Documented intended/out-of-scope use in model card + EU AI Act classification

A.10 — Third parties & customers

Ctrl Control Status How / artifact
A.10.2 Allocation of responsibilities A Offline default has no third party; cloud seam = deployer-of-foundation-model, responsibility documented in risk register R5
A.10.3 Suppliers A Foundation-model provider (Bedrock/Anthropic) only in opt-in seam; fail-closed + circuit breaker; deps pinned + SBOM
A.10.4 Customer expectations N/A No paying customers / no contractual deployment; open-source reference implementation. (Re-applies if offered as a managed service.)

Summary

Of the Annex A controls, the operative set (A.2–A.10) is applicable and implemented; two items are justified deferrals tied to a documented trigger: the full AI System Impact Assessment (A.5.3 / Cl. 6.1.4) and A.10.4 customer expectations, both of which activate when household-data personalization or a managed-service offering ships. No control is unaddressed by silence.

The management-system claim Sprout makes: the controls that matter most for an AI assistant — provenance (A.7.5), V&V (A.6.2.4), responsible-use objectives (A.9.3) — are not policy documents but code paths in src/sprout/, verified every run by the eval harness and gated by make verify.

Cross-references