Standards alignment

September 5, 2026 · View on GitHub

What this is: a traceable self-assessment — for each clause of the AI-testing / AI-quality standards and each ISO/IEC 42001 Annex A control considered here (Annex A controls are reference controls chosen through applicability and risk treatment, the way a Statement of Applicability would — they are not automatic requirements), the FH mechanism that answers it and the file where the evidence lives. What it is not: a conformity assessment or a certification. Half of the ISO/IEC 42119 series is still a draft Technical Specification (DTS / AWI), so the word used here is alignment, never compliance. Full crosswalk with sources: knowledge/shared/harness-core/iso_ai_standards_crosswalk.md.

Why the harness is the right unit

The standards' "AI system" is the model plus what runs it — prompts, gates, logs, intervention points, risk records. The clauses a model vendor cannot answer for you (42001 A.6.2.4 verification & validation, A.6.2.8 event logs, 25059 intervenability, TS 8200 controllability, 42119-2 §6 risk identification) live exactly in the harness layer. FH is a meta-harness, so the same table applies to FH itself and to the field harnesses it emits.

Summary table

Standard · clauseRequirement (gist)FH mechanismEvidenceStatus
ISO/IEC TS 42119-2:2025 §5.4/§6risk-based testing, risk identification6-axis verification chosen by failure mode; marker defeater: / affected:.axes_23_passed_*.marker🟡 spread over commit-level fields, no risk register
42119-2 §5.5 (29119-2 dynamic test process)design → implement → execute → reportlane suites → selfcheck.sh → CI validate · 4-axis pre-commit gatetemplates/.git-hooks/pre-commit, scripts/selfcheck.sh
42119-2 §5.5 (29119-2 organizational / management processes)policy · strategy · planning · monitoring & control · completionpolicy = CLAUDE.md gate sections · plan = pre-registration seal · completion = marker — no written monitoring/control or completion-criteria processCLAUDE.md, PREREG_*.md🟡 activities exist, process documents do not
42119-2 §5.6 (29119-3 documentation)plan · case/procedure spec · completion report · incident reportpre-registration seal · lane files · marker + PR body · fh_signal_*scripts/test_*_lanes.sh, markers🟡 1:1 but not named by 29119-3 template names — this page is the name tag
42119-2 §7.2test levels unit → integration → system → acceptancelanes → selfcheck → isolated-clone sim; first real use is a field trial, not acceptance testing (no acceptance criteria or sign-off evidence)scripts/sim_isolated_run.sh🟡 unit–system ✅ · acceptance ❌
42119-2 §7.3.4 / TR 29119-11 §8black-box model-testing techniques: A/B, back-to-back, adversarial, metamorphic (drift is a monitoring / model-update concern in TR 29119-11; "drift testing" as a technique name comes from the ISTQB CT-AI v2 syllabus)ARM/CTRL one-variable sim (A/B) · cross-family review (back-to-back) · challenger / steel-quench (adversarial) · nightly live-eval (drift monitoring) · revert probes (mutation)probes_live.yaml, probe_live_eval.sh🟡 metamorphic relations not written as such
TR 29119-11 test-oracle problemexpected results when none existknown-pair controls in the same run · scorer fixed before results · "not found ≠ 0" · UNCALIBRATEDmeasurement-integrity-checklist.md✅ mechanism · ✅ oracle type recorded — marker field oracle: (closed enum: known-pair · metamorphic · back-to-back · a-b · human · none; none needs its reason; form-checked by the pre-commit hook, 2026-09-05)
TR 29119-11 non-determinismstatistical treatmentreps>=3 bar · 3/3 or 0/3 or no separation · controlsprobe_scope_check.sh
DTS 42119-3.2 (stage 50.20)V&V analysis: simulation, evaluation, formal methodsisolated-clone simulation · live-eval thresholds · no formal methods🟡
AWI TS 42119-7 red teaming (public scope only — the topics it covers, not normative requirements)terms · risks · applicability · objectives & attack vectors · method · reporting · life-cycle integrationcross-family adversarial panel (a reviewer that never saw the author's reasoning) · attack-angle registry · findings typed S/A/B · mandatory before commit on load-bearing changesauto-decorrelation SKILL, field_verdict_crossfamily_gate.md, templates/RED_TEAM_REPORT.md🟡 → report template added 2026-09-05
AWI TS 42119-8 (stage 20.00)quality assessment of prompt-based text-to-text GenAIlive probes with expectation/control regex and polarity · nightly isolated run · tier × effort expectation doc.claude/regression/probes_live.yaml, docs/model_tier_expectations.md🟡
ISO/IEC 25059 robustnesskeep working under adversarial / invalid inpute.g. padding-bypass fix on the outbound hook (10 KB: 47 s → 0.23 s) with truncation / string / key fail-closedscripts/test_outbound_query_hook_lanes.sh H23–H30✅ by case
25059 user controllability · intervenability · TS 8200timely human intervention, control pointsHITL gates · consent leases (quoted words · dated · scoped) · autonomy floor · explicit, logged override channels (DESTRUCTIVE_OP_OK, PUBLIC_SURFACE_OK, PUSH_ZONE_OK)CLAUDE.md §Agent Dispatch, templates/.git-hooks/pre-push✅ strong — every control point is a logged channel
25059 transparencyappropriate information to stakeholdersmarker "axes run and control liveness" · PR body "residuals named" · wiki status board "what proved it"docs/OUTPUT_EVIDENCE.md
25059 (DIS) service traceabilitydefinition differs between public secondary sources (logging of models / datasets / requests / outcomes vs. traceability of service outcomes to user needs) — both readings kept until the text is readlogging reading: governance log · hook event files (counts and labels, never values) · sub-agent ledger · sim header (corpus_head_date, sim_model, sim_model_cutoff) · user-need reading: marker soul:defeater: ↔ lanestracks/_meta/governance_log_*.yaml, knowledge/shared/learnings/subagent_invocations_log.yaml🟡 strong on the logging reading (which is also the 42001 A.6.2.8 evidence)
ISO/IEC 42001 A.6.2.4 V&Vverify against defined criteria4-axis gate: required fields, closed enums, non-vacuous groundspre-commit validate_*_leg()
42001 A.6.2.6 operation & monitoringmonitoring with criteria, review, response, ownershipmechanisms: nightly live-eval · daily digest · daily report (launchd) · "verification env ≠ runtime env" discipline; thresholds decided after a calibration weekscripts/com.forge-harness.*.plist🟡 mechanisms exist, no written monitoring procedure
42001 A.6.2.7 / A.8technical documentation · stakeholder informationknowledge/, docs/USER_GUIDE.md, docs/USE_CASES.md, this page
42001 A.9 (A.9.2 responsible-use processes · A.9.3 objectives · A.9.4 intended use)responsible useintended use is communicated (guides); use processes = consent leases, autonomy floor, irreversible-surface gates; no written responsible-use objectivesCLAUDE.md §Agent Dispatch🟡
42001 A.6.2.8 event logslog eventssee traceability row
42001 A.5 impact assessmentassess impactschamber §3-SCREEN screens KILL / NOT-APPLICABLE / CURATED / EMIT — does not ask "impact on whom"harness_incubator_doctrine.md🟡
42001 A.7 data (A.7.2–A.7.6)acquisition, quality, provenance, preparationFH holds no training data but does hold evaluation data (probes, corpora, transcripts, logs): provenance = sim header corpus_head_date · quality = known-pair calibration · preparation = residency strip · company residency is absoluteprobes_live.yaml, scripts/residency_closure_scan.py🟡 provenance/preparation present; no written acquisition or quality criteria for evaluation data
ISO/IEC 5338 life cycledefine · control · execute · improveincubator (chamber) → EMIT → field → harvest-loop compoundingharness_incubator_doctrine.md, hub_compounding_loop.md
ISO/IEC 23894 risk managementidentify · analyse · evaluate · treatship-readiness grades · Surface-Class Degrade Invariant (irreversible = fail-closed)ship_readiness_gate.md, CLAUDE.md §Irreversibility Gates✅ partial
ISO/IEC 20246 reviewswork-product reviewscross-family review · /apex-review · hub-cc-pr-reviewer · cold read (beginner)plugins/fh-meta/agents/

What a field harness inherits

A harness FH emits gets the same rows for free where it adopts the templates (templates/.git-hooks/*, the marker schema, the red-team report). For a QA harness the primary target is ISO/IEC/IEEE 29119 itself: test design & implementation (act 1), work-product review (act 1.5, ISO/IEC 20246), execution & incident reporting (act 2), completion & regression (act 3). That mapping lives in the crosswalk §3.

Honest boundaries

  • Read from public material: the 42119-2 sample (table of contents, foreword), scope statements of 42119-3.2/7/8, secondary summaries, and the ISTQB CT-AI v2.0 syllabus as a public training syllabus that covers the same AI-testing topics (it is not a mirror of TR 29119-11 — v2 extends into GenAI and red teaming). Clause bodies of the paid texts were not read; the DIS 25059 definition of service traceability differs between the two secondary sources consulted and is kept as two readings.
  • Not covered here (named so the map is not read as complete): ISO/IEC TS 25058, TS 6254 / 12792, 24029-2/3, TS 24970 (AI logging — the first candidate for the next revision), NIST AI RMF, IEEE 7001, EU AI Act Art. 9/15.
  • Self-assessment, so gameable in the same way every self-declared field is (CLAUDE.md §자기 대조). The cross-family review on the commit that added this page is the only independent check applied so far.
  • One author, one repository, one day.