"""squash/audit_sim.py — Sprint 22 W229–W231 (Track C / C5). Regulatory Examination Simulation Engine: ``squash simulate-audit``. The premise: 78% of business executives lack confidence they could pass an independent AI governance audit within 90 days (Grant Thornton 2026). This module runs a mock examination *from the regulator's perspective*, pulls answers from the model's squash attestation artefacts, flags gaps where evidence is missing, and produces an executive-ready readiness report — the $5K–$15K professional-service deliverable compressed into a 60-second CLI command. Architecture ============ ``ExamQuestion`` A single examiner question: what article it comes from, what squash artefact files would answer it, which CLI commands generate that evidence, and how critical it is (weight 1–3; weight-3 = critical gate). ``ExamAnswer`` The engine's answer for one question: PASS / PARTIAL / FAIL, evidence found/missing, gap description, remediation step, and a squash command the org can run *right now* to close the gap. ``ReadinessReport`` Aggregate: overall score 0–100, readiness tier, critical-gate summary, prioritised remediation roadmap, executive summary, and renders to JSON + Markdown. ``AuditSimulator`` Stateless. ``.simulate(model_path, regulator)`` → ``ReadinessReport``. Evidence detection is file-presence based (no network calls) so it runs instantly in air-gapped CI. Regulator profiles ================== ``EU-AI-Act`` 38 questions covering Art. 9–15, 17, 73, Annex IV, and conformity assessment for high-risk AI. ``NIST-RMF`` 30 questions spanning GOVERN, MAP, MEASURE, MANAGE. ``SEC`` 22 questions on AI disclosure, material risk, ops controls, and investment-adviser obligations (OMB M-26-04 ready). ``FDA`` 20 questions on SaMD risk, clinical validation, post-market, bias/fairness, and device labelling. Scoring ======= Each question contributes ``question.weight × points_for_status`` to the total score, where ``points_for_status`` = 2 (PASS) / 1 (PARTIAL) / 0 (FAIL). ``overall_score = 100 × Σ(earned) / Σ(max_possible)`` **Critical-gate cap:** any weight-3 question that scores FAIL caps the total at 74 (just below "Substantial") regardless of other scores. Tiers:: AUDIT_READY ≥ 80 · no critical fails SUBSTANTIAL 60–79 (or 74 cap from critical fail) DEVELOPING 40–59 EARLY_STAGE < 40 Stdlib-only. No external dependencies. Evidence detection is pure filesystem introspection against a local model directory. """ from __future__ import annotations import datetime import json import logging from dataclasses import dataclass, field from pathlib import Path from typing import Any log = logging.getLogger(__name__) # ── Known squash artefact filenames ─────────────────────────────────────────── # Canonical squash artefact names — searched inside model_path/ and model_path/squash/ _ATTEST = "squash-attest.json" _ATTEST_ALT = "squash_attestation.json" _BOM = "cyclonedx-mlbom.json" _SPDX = "spdx-mlbom.json" _SCAN = "squash-scan.json" _VEX = "squash-vex-report.json" _ANNEX_IV = "annex_iv.json" _MODEL_CARD = "squash-model-card-hf.md" _BIAS = "bias_audit_report.json" _LINEAGE = "data_lineage_certificate.json" _INCIDENT = "squash-incident.json" _DRIFT = "squash-drift.json" _NIST_RMF = "nist_rmf_report.json" _SQUASH_CFG = ".squash.yml" _SQUASH_JSON = "squish.json" _POLICY_PREFIX = "squash-policy-" # matches squash-policy-eu-ai-act.json etc. _CHAIN_ATTEST = "chain-attest.json" _REG_GATE = "registry-gate.json" _SBOM_DIFF = "squash-sbom-diff.json" _ALL_KNOWN = frozenset({ _ATTEST, _ATTEST_ALT, _BOM, _SPDX, _SCAN, _VEX, _ANNEX_IV, _MODEL_CARD, _BIAS, _LINEAGE, _INCIDENT, _DRIFT, _NIST_RMF, _SQUASH_CFG, _SQUASH_JSON, _CHAIN_ATTEST, _REG_GATE, _SBOM_DIFF, }) # ── Core data classes ───────────────────────────────────────────────────────── @dataclass class ExamQuestion: """A single examiner question with evidence pointers. Attributes: q_id: Unique question ID (e.g. ``"EU-001"``). article: Regulatory article/section (e.g. ``"Art. 9(1)a"``). question: The examiner's question text. answer_sources: Squash artefact filenames that answer this question. Any one present → PARTIAL; all present → PASS. answer_cli: Squash CLI commands that generate the missing evidence. weight: 1 = standard · 2 = important · 3 = critical gate. category: Thematic category (e.g. ``"risk-management"``). days_to_close: Estimated engineer-days to produce evidence. """ q_id: str article: str question: str answer_sources: list[str] answer_cli: list[str] weight: int = 2 category: str = "general" days_to_close: int = 3 @dataclass class ExamAnswer: """Engine answer for one examiner question.""" question: ExamQuestion status: str # "PASS" | "PARTIAL" | "FAIL" | "N/A" evidence_found: list[str] = field(default_factory=list) evidence_missing: list[str] = field(default_factory=list) gap_description: str = "" remediation: str = "" def to_dict(self) -> dict[str, Any]: return { "q_id": self.question.q_id, "article": self.question.article, "question": self.question.question, "category": self.question.category, "weight": self.question.weight, "status": self.status, "evidence_found": list(self.evidence_found), "evidence_missing": list(self.evidence_missing), "gap_description": self.gap_description, "remediation": self.remediation, "squash_commands": list(self.question.answer_cli), "days_to_close": self.question.days_to_close, } @dataclass class ReadinessReport: """Aggregate examination result with score, tier, and remediation plan.""" regulator: str model_path: str generated_at: str overall_score: int readiness_tier: str # AUDIT_READY|SUBSTANTIAL|DEVELOPING|EARLY_STAGE answers: list[ExamAnswer] critical_fails: int total_questions: int passing: int partial: int failing: int high_priority_gaps: list[str] = field(default_factory=list) executive_summary: str = "" squash_version: str = "audit_sim_v1" # ── Serialisation ───────────────────────────────────────────────────── def to_dict(self) -> dict[str, Any]: return { "squash_version": self.squash_version, "regulator": self.regulator, "model_path": self.model_path, "generated_at": self.generated_at, "overall_score": self.overall_score, "readiness_tier": self.readiness_tier, "summary": { "total_questions": self.total_questions, "passing": self.passing, "partial": self.partial, "failing": self.failing, "critical_fails": self.critical_fails, }, "high_priority_gaps": list(self.high_priority_gaps), "executive_summary": self.executive_summary, "answers": [a.to_dict() for a in self.answers], "remediation_roadmap": self._roadmap_dicts(), } def to_json(self, indent: int = 2) -> str: return json.dumps(self.to_dict(), indent=indent, sort_keys=True) def to_markdown(self) -> str: tier_emoji = { "AUDIT_READY": "✅", "SUBSTANTIAL": "🟡", "DEVELOPING": "🟠", "EARLY_STAGE": "🔴", }.get(self.readiness_tier, "⚪") lines: list[str] = [ f"# {self.regulator} Regulatory Examination Simulation", "", f"**Model:** `{self.model_path}` ", f"**Generated:** {self.generated_at[:10]} ", f"**Readiness:** {tier_emoji} **{self.readiness_tier.replace('_', ' ')}** ", f"**Score:** {self.overall_score}/100", "", "---", "", "## Executive Summary", "", self.executive_summary, "", "## Scorecard", "", f"| Metric | Value |", f"|---|---|", f"| Overall score | **{self.overall_score}**/100 |", f"| Readiness tier | {self.readiness_tier.replace('_', ' ')} |", f"| Questions answered (PASS) | {self.passing}/{self.total_questions} |", f"| Partially answered | {self.partial} |", f"| Gaps (FAIL) | {self.failing} |", f"| Critical gate fails | {self.critical_fails} |", "", ] # Category breakdown categories: dict[str, list[ExamAnswer]] = {} for ans in self.answers: cat = ans.question.category categories.setdefault(cat, []).append(ans) lines.extend(["## Results by Category", ""]) for cat, answers in sorted(categories.items()): passing = sum(1 for a in answers if a.status == "PASS") total = len(answers) pct = int(100 * passing / total) bar = "█" * (pct // 10) + "░" * (10 - pct // 10) lines.append( f"**{cat.replace('-', ' ').title()}** {bar} {pct}% " f"({passing}/{total})" ) lines.extend(["", "## Question Detail", ""]) for ans in self.answers: icon = {"PASS": "✅", "PARTIAL": "🟡", "FAIL": "❌", "N/A": "⊘"}.get( ans.status, "?" ) weight_star = "★" * ans.question.weight lines.extend([ f"### {ans.question.q_id} — {icon} {ans.question.article}", "", f"**{ans.question.question}**", "", f"*Weight: {weight_star} · Category: {ans.question.category}*", "", ]) if ans.evidence_found: lines.append("Evidence found:") for ev in ans.evidence_found: lines.append(f" - `{ev}`") if ans.evidence_missing: lines.append("Evidence missing:") for ev in ans.evidence_missing: lines.append(f" - `{ev}`") if ans.gap_description: lines.extend(["", f"> **Gap:** {ans.gap_description}"]) if ans.remediation and ans.status != "PASS": lines.extend(["", f"**Remediation:** {ans.remediation}"]) if ans.question.answer_cli: lines.append("") lines.append("Run:") for cmd in ans.question.answer_cli[:2]: lines.append(f"```bash\n{cmd}\n```") lines.append("") # Remediation roadmap roadmap = self._roadmap() if roadmap: lines.extend([ "---", "", "## 90-Day Remediation Roadmap", "", "Sequenced to maximise compliance score per engineering-day invested.", "", "| Priority | Gap | Squash command | Days |", "|---|---|---|---|", ]) day_counter = 0 for item in roadmap: day_counter += item["days_to_close"] lines.append( f"| **{item['priority']}** | {item['article']}: " f"{item['q_id']} | `{item['command']}` | {item['days_to_close']} |" ) lines.extend([ "", f"**Estimated total remediation effort: {day_counter} engineer-days**", "", ]) lines.extend([ "---", "", "*Generated by [squash](https://getsquash.dev) · " f"`squash simulate-audit --regulator {self.regulator}` · " "Squash violations, not velocity.*", ]) return "\n".join(lines) + "\n" def save( self, output_dir: Path | str, stem: str = "audit-readiness", ) -> dict[str, Path]: output_dir = Path(output_dir) output_dir.mkdir(parents=True, exist_ok=True) written: dict[str, Path] = {} p = output_dir / f"{stem}.json" p.write_text(self.to_json(), encoding="utf-8") written["json"] = p p = output_dir / f"{stem}.md" p.write_text(self.to_markdown(), encoding="utf-8") written["md"] = p return written # ── Private ─────────────────────────────────────────────────────────── def _roadmap(self) -> list[dict[str, Any]]: """Prioritised remediation list for failed/partial answers.""" items: list[dict[str, Any]] = [] for ans in self.answers: if ans.status in ("FAIL", "PARTIAL"): cmd = ans.question.answer_cli[0] if ans.question.answer_cli else "squash attest" items.append({ "q_id": ans.question.q_id, "article": ans.question.article, "weight": ans.question.weight, "status": ans.status, "command": cmd, "days_to_close": ans.question.days_to_close, "priority": ( "CRITICAL" if ans.question.weight == 3 else "HIGH" if ans.question.weight == 2 else "MEDIUM" ), }) # Sort: critical first, then weight desc, then days asc (quick wins) items.sort(key=lambda x: (-x["weight"], x["days_to_close"])) return items def _roadmap_dicts(self) -> list[dict[str, Any]]: return self._roadmap() # ── Simulation engine ───────────────────────────────────────────────────────── class AuditSimulator: """Stateless regulatory examination simulator. Usage:: report = AuditSimulator().simulate( model_path=Path("./my-model"), regulator="EU-AI-Act", ) report.save("./out") print(report.to_markdown()) """ SUPPORTED_REGULATORS: frozenset[str] = frozenset({ "EU-AI-Act", "NIST-RMF", "SEC", "FDA", }) def simulate( self, model_path: Path | str, regulator: str, ) -> ReadinessReport: """Run the mock examination and return a ``ReadinessReport``. Args: model_path: Path to a model directory containing squash artefacts. regulator: One of ``"EU-AI-Act"``, ``"NIST-RMF"``, ``"SEC"``, ``"FDA"``. Raises: ValueError: If ``regulator`` is not supported. """ model_path = Path(model_path) reg = regulator.upper().replace(" ", "-") # Normalise aliases _aliases = {"EU": "EU-AI-ACT", "NIST": "NIST-RMF", "EU-AI-ACT": "EU-AI-ACT"} reg = _aliases.get(reg, reg) profile_fn = { "EU-AI-ACT": _eu_ai_act_questions, "EU-AI-Act".upper().replace(" ", "-"): _eu_ai_act_questions, "NIST-RMF": _nist_rmf_questions, "SEC": _sec_questions, "FDA": _fda_questions, }.get(reg.upper()) if profile_fn is None: raise ValueError( f"Unsupported regulator: {regulator!r}. " f"Supported: {sorted(self.SUPPORTED_REGULATORS)}" ) questions = profile_fn() artifacts = _collect_artifacts(model_path) answers = [_answer_question(q, artifacts, model_path) for q in questions] score, tier, critical_fails = _compute_score(answers) passing = sum(1 for a in answers if a.status == "PASS") partial = sum(1 for a in answers if a.status == "PARTIAL") failing = sum(1 for a in answers if a.status == "FAIL") high_gaps = [ f"{a.question.q_id} ({a.question.article}): {a.gap_description}" for a in answers if a.status == "FAIL" and a.question.weight == 3 ] summary = _executive_summary(regulator, score, tier, passing, partial, failing, critical_fails, len(questions)) return ReadinessReport( regulator=regulator, model_path=str(model_path), generated_at=datetime.datetime.now(datetime.timezone.utc).isoformat(), overall_score=score, readiness_tier=tier, answers=answers, critical_fails=critical_fails, total_questions=len(questions), passing=passing, partial=partial, failing=failing, high_priority_gaps=high_gaps, executive_summary=summary, ) # ── Evidence detection ──────────────────────────────────────────────────────── def _collect_artifacts(model_path: Path) -> set[str]: """Return the set of squash artefact filenames present in model_path.""" present: set[str] = set() if not model_path.exists(): return present for search_dir in (model_path, model_path / "squash"): if not search_dir.is_dir(): continue for child in search_dir.iterdir(): nm = child.name if nm in _ALL_KNOWN or nm.startswith(_POLICY_PREFIX): present.add(nm) return present def _answer_question( q: ExamQuestion, artifacts: set[str], model_path: Path, ) -> ExamAnswer: """Score one examiner question against the available evidence.""" required = list(q.answer_sources) found = [s for s in required if s in artifacts] missing = [s for s in required if s not in artifacts] # Special case: policy prefix match for s in list(missing): if s == _POLICY_PREFIX: if any(a.startswith(_POLICY_PREFIX) for a in artifacts): missing.remove(s) found.append(s) ratio = len(found) / max(len(required), 1) if not required: status = "N/A" gap = "" remediation = "" elif ratio >= 0.85: status = "PASS" gap = "" remediation = "" elif ratio > 0: status = "PARTIAL" gap = ( f"Some evidence present but {len(missing)} artefact(s) missing: " f"{', '.join(missing[:3])}." ) cmd = q.answer_cli[0] if q.answer_cli else "squash attest" remediation = f"Generate missing evidence: `{cmd}`" else: status = "FAIL" gap = ( f"No evidence found. Required artefact(s): " f"{', '.join(missing[:3])}." ) cmd = q.answer_cli[0] if q.answer_cli else "squash attest" remediation = ( f"No squash artefacts found for this requirement. " f"Start with: `{cmd}`" ) return ExamAnswer( question=q, status=status, evidence_found=found, evidence_missing=missing, gap_description=gap, remediation=remediation, ) def _compute_score( answers: list[ExamAnswer], ) -> tuple[int, str, int]: """Return (overall_score_0_100, readiness_tier, critical_fail_count).""" _POINTS = {"PASS": 2, "PARTIAL": 1, "FAIL": 0, "N/A": 0} earned = sum(a.question.weight * _POINTS[a.status] for a in answers) max_pts = sum(a.question.weight * 2 for a in answers if a.status != "N/A") raw = int(100 * earned / max(max_pts, 1)) critical_fails = sum( 1 for a in answers if a.question.weight == 3 and a.status == "FAIL" ) score = min(raw, 74) if critical_fails > 0 else raw score = max(0, min(100, score)) if score >= 80 and critical_fails == 0: tier = "AUDIT_READY" elif score >= 60: tier = "SUBSTANTIAL" elif score >= 40: tier = "DEVELOPING" else: tier = "EARLY_STAGE" return score, tier, critical_fails def _executive_summary( regulator: str, score: int, tier: str, passing: int, partial: int, failing: int, critical_fails: int, total: int, ) -> str: tier_desc = { "AUDIT_READY": "examination-ready with strong evidence coverage", "SUBSTANTIAL": "substantially prepared, with targeted gaps to close", "DEVELOPING": "in active development of compliance evidence", "EARLY_STAGE": "at an early stage of compliance evidence generation", }.get(tier, "assessed") crit_note = ( f" **{critical_fails} critical-gate requirement(s) are missing** — " "these must be addressed before examination regardless of overall score." if critical_fails else " No critical-gate requirements are missing." ) gap_sentence = ( f"Of {total} examiner questions, **{passing} answered** (PASS), " f"**{partial} partially answered** (PARTIAL), and " f"**{failing} unanswered** (FAIL)." ) return ( f"This model portfolio is **{tier_desc}** for a {regulator} regulatory " f"examination, with an overall readiness score of **{score}/100**. " f"{gap_sentence}{crit_note} " f"The remediation roadmap below is sequenced by weight and effort — " f"following it will produce audit-ready evidence in the shortest elapsed time." ) # ── Regulator profiles ──────────────────────────────────────────────────────── # Each function returns the ordered list of examiner questions. # Questions are ordered: critical gates first, then high-weight, then standard. # Reference source for EU AI Act: Regulation (EU) 2024/1689. # Reference source for NIST RMF: NIST AI 100-1. # Reference source for SEC: OMB M-26-04, SEC AI priorities 2026. # Reference source for FDA: FDA AI/ML-Based SaMD Action Plan. def _eu_ai_act_questions() -> list[ExamQuestion]: """38 examiner questions for EU AI Act (Regulation (EU) 2024/1689).""" return [ # ── Critical gates (weight 3) ───────────────────────────────────── ExamQuestion( q_id="EU-001", article="Art. 9(1)", weight=3, category="risk-management", question="Has the organisation established and maintained a risk management system for this AI system throughout its lifecycle?", answer_sources=[_ATTEST, _SCAN, "squash-policy-eu-ai-act.json"], answer_cli=["squash attest --policy eu-ai-act", "squash risk-assess"], days_to_close=2, ), ExamQuestion( q_id="EU-002", article="Art. 11(1)", weight=3, category="technical-documentation", question="Is the technical documentation (Annex IV) complete and current, covering all 12 sections?", answer_sources=[_ANNEX_IV, _BOM], answer_cli=["squash annex-iv generate --format all", "squash attest"], days_to_close=3, ), ExamQuestion( q_id="EU-003", article="Art. 14(1)", weight=3, category="human-oversight", question="Are human oversight measures implemented and documented for this AI system?", answer_sources=[_ANNEX_IV, _ATTEST], answer_cli=["squash annex-iv generate", "squash attest --policy eu-ai-act"], days_to_close=3, ), ExamQuestion( q_id="EU-004", article="Art. 13(1)", weight=3, category="transparency", question="Has the AI system been designed to enable transparency and interpretability for deployers?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format all", "squash annex-iv generate"], days_to_close=2, ), ExamQuestion( q_id="EU-005", article="Art. 73(1)", weight=3, category="incident-reporting", question="Are procedures in place to report serious incidents to market surveillance authorities within 15 working days?", answer_sources=[_INCIDENT, _ATTEST], answer_cli=["squash incident --type serious", "squash freeze --help"], days_to_close=5, ), # ── High weight (weight 2) ──────────────────────────────────────── ExamQuestion( q_id="EU-006", article="Art. 10(2)", weight=2, category="data-governance", question="Have training, validation, and testing data sets been governed with documented provenance and quality criteria?", answer_sources=[_LINEAGE, _ANNEX_IV], answer_cli=["squash data-lineage --model ./model", "squash annex-iv generate"], days_to_close=3, ), ExamQuestion( q_id="EU-007", article="Art. 10(5)", weight=2, category="data-governance", question="Have the datasets been assessed for potential biases that could produce discriminatory outputs?", answer_sources=[_BIAS, _ANNEX_IV], answer_cli=["squash bias-audit --model ./model", "squash annex-iv generate"], days_to_close=5, ), ExamQuestion( q_id="EU-008", article="Art. 12(1)", weight=2, category="record-keeping", question="Are logs automatically generated and retained to enable post-market monitoring?", answer_sources=[_ATTEST, _SCAN, _DRIFT], answer_cli=["squash attest --sign", "squash drift-check"], days_to_close=2, ), ExamQuestion( q_id="EU-009", article="Art. 15(1)", weight=2, category="accuracy-robustness", question="Has the AI system been validated for accuracy, robustness, and cybersecurity appropriate to its intended purpose?", answer_sources=[_SCAN, _VEX, "squash-policy-eu-ai-act.json"], answer_cli=["squash scan ./model", "squash vex update", "squash attest --policy eu-ai-act"], days_to_close=3, ), ExamQuestion( q_id="EU-010", article="Art. 15(3)", weight=2, category="accuracy-robustness", question="Are measures in place to protect against adversarial attacks and model poisoning?", answer_sources=[_SCAN, _VEX], answer_cli=["squash scan ./model --exit-2-on-unsafe", "squash vex update"], days_to_close=3, ), ExamQuestion( q_id="EU-011", article="Art. 9(4)", weight=2, category="risk-management", question="Have residual risks been evaluated and documented after risk mitigation measures?", answer_sources=["squash-policy-eu-ai-act.json", _ANNEX_IV], answer_cli=["squash risk-assess", "squash annex-iv generate"], days_to_close=2, ), ExamQuestion( q_id="EU-012", article="Art. 9(9)", weight=2, category="risk-management", question="Is the risk management system subject to continuous updates based on post-market monitoring data?", answer_sources=[_DRIFT, _ATTEST], answer_cli=["squash drift-check --baseline ./baseline.json", "squash watch"], days_to_close=2, ), ExamQuestion( q_id="EU-013", article="Art. 10(3)", weight=2, category="data-governance", question="Have data preparation pipelines been documented including pre-processing steps?", answer_sources=[_LINEAGE, _BOM], answer_cli=["squash data-lineage --model ./model", "squash attest"], days_to_close=3, ), ExamQuestion( q_id="EU-014", article="Art. 16(a)", weight=2, category="conformity", question="Has the provider established and maintained a quality management system (QMS) for this AI system?", answer_sources=[_ATTEST, _SQUASH_CFG], answer_cli=["squash init", "squash attest --policy eu-ai-act"], days_to_close=5, ), ExamQuestion( q_id="EU-015", article="Art. 17(1)", weight=2, category="conformity", question="Is there a documented quality management system covering design, development, testing, and deployment?", answer_sources=[_ANNEX_IV, _ATTEST, _SQUASH_CFG], answer_cli=["squash annex-iv generate", "squash init"], days_to_close=5, ), ExamQuestion( q_id="EU-016", article="Art. 13(3)(b)", weight=2, category="transparency", question="Is the model card accurate, current, and describing the AI system's intended purpose and capabilities?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format hf --validate", "squash annex-iv generate"], days_to_close=2, ), ExamQuestion( q_id="EU-017", article="Art. 61(1)", weight=2, category="post-market", question="Is a post-market monitoring plan implemented, including systematic collection of performance data?", answer_sources=[_DRIFT, _ATTEST, _SCAN], answer_cli=["squash drift-check", "squash watch", "squash monitor"], days_to_close=3, ), ExamQuestion( q_id="EU-018", article="Art. 26(1)", weight=2, category="deployer-obligations", question="Have deployer obligations been documented, including use-case conformity and human oversight measures?", answer_sources=[_ANNEX_IV, _MODEL_CARD], answer_cli=["squash annex-iv generate", "squash model-card --format eu-ai-act"], days_to_close=2, ), ExamQuestion( q_id="EU-019", article="Annex IV §1", weight=2, category="technical-documentation", question="Does Annex IV §1 describe the AI system's general purpose and intended use with sufficient detail?", answer_sources=[_ANNEX_IV], answer_cli=["squash annex-iv generate --system-name 'System Name' --intended-purpose '...'"], days_to_close=1, ), ExamQuestion( q_id="EU-020", article="Annex IV §2", weight=2, category="technical-documentation", question="Are training data, datasets, and data processing methods documented in Annex IV §2?", answer_sources=[_ANNEX_IV, _LINEAGE], answer_cli=["squash annex-iv generate", "squash data-lineage"], days_to_close=2, ), ExamQuestion( q_id="EU-021", article="Annex IV §3", weight=2, category="technical-documentation", question="Is the system architecture, including hardware requirements and model components, documented?", answer_sources=[_ANNEX_IV, _BOM], answer_cli=["squash annex-iv generate", "squash attest"], days_to_close=1, ), ExamQuestion( q_id="EU-022", article="Annex IV §6", weight=2, category="technical-documentation", question="Are performance metrics, validation results, and benchmark results documented?", answer_sources=[_ANNEX_IV, _ATTEST], answer_cli=["squash annex-iv generate", "squash attest --policy eu-ai-act"], days_to_close=2, ), ExamQuestion( q_id="EU-023", article="Art. 9(6)", weight=2, category="risk-management", question="Have tests been conducted to identify risk management measures effective for the AI system?", answer_sources=[_SCAN, "squash-policy-eu-ai-act.json", _ATTEST], answer_cli=["squash attest --policy eu-ai-act", "squash scan ./model"], days_to_close=2, ), ExamQuestion( q_id="EU-024", article="Art. 72(1)", weight=2, category="incident-reporting", question="Are logs and incident reports maintained for market surveillance purposes?", answer_sources=[_INCIDENT, _ATTEST, _SCAN], answer_cli=["squash incident", "squash attest --sign"], days_to_close=3, ), # ── Standard (weight 1) ─────────────────────────────────────────── ExamQuestion( q_id="EU-025", article="Art. 10(6)", weight=1, category="data-governance", question="Have measures been taken to detect and address dataset biases across protected attributes?", answer_sources=[_BIAS], answer_cli=["squash bias-audit --model ./model"], days_to_close=5, ), ExamQuestion( q_id="EU-026", article="Art. 14(4)", weight=1, category="human-oversight", question="Do human oversight measures include the ability to halt the AI system in case of anomalous behaviour?", answer_sources=[_INCIDENT, _SQUASH_CFG], answer_cli=["squash freeze --help", "squash incident"], days_to_close=5, ), ExamQuestion( q_id="EU-027", article="Art. 25(1)", weight=1, category="supply-chain", question="Have agreements with distributors and importers of this AI system been documented?", answer_sources=[_ATTEST, _BOM], answer_cli=["squash attest --sign", "squash trust-package"], days_to_close=5, ), ExamQuestion( q_id="EU-028", article="Art. 53(1)", weight=1, category="gpai", question="If this is a general-purpose AI model, has technical documentation per Art. 53 been prepared?", answer_sources=[_ANNEX_IV, _BOM, _MODEL_CARD], answer_cli=["squash annex-iv generate --format all", "squash model-card --format all"], days_to_close=3, ), ExamQuestion( q_id="EU-029", article="Art. 51(1)", weight=1, category="gpai", question="Has a risk classification been performed for the AI system with documented justification?", answer_sources=["squash-policy-eu-ai-act.json", _ANNEX_IV], answer_cli=["squash attest --policy eu-ai-act", "squash risk-assess"], days_to_close=2, ), ExamQuestion( q_id="EU-030", article="Annex IV §4", weight=1, category="technical-documentation", question="Are deployment, operational, and maintenance requirements documented?", answer_sources=[_ANNEX_IV], answer_cli=["squash annex-iv generate"], days_to_close=1, ), ExamQuestion( q_id="EU-031", article="Annex IV §5", weight=1, category="technical-documentation", question="Is the risk management process and its outcomes documented per Annex IV §5?", answer_sources=[_ANNEX_IV, "squash-policy-eu-ai-act.json"], answer_cli=["squash annex-iv generate", "squash attest --policy eu-ai-act"], days_to_close=1, ), ExamQuestion( q_id="EU-032", article="Art. 50(1)", weight=1, category="transparency", question="Are AI-generated content disclosure mechanisms implemented when required?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format eu-ai-act"], days_to_close=3, ), ExamQuestion( q_id="EU-033", article="Art. 9(8)", weight=1, category="risk-management", question="Are risk management records kept for at least 10 years after the AI system is placed on the market?", answer_sources=[_ATTEST, _ANNEX_IV], answer_cli=["squash attest --sign", "squash annual-review"], days_to_close=2, ), ExamQuestion( q_id="EU-034", article="Art. 16(d)", weight=1, category="conformity", question="Is there a Declaration of Conformity prepared and signed by the provider?", answer_sources=[_ATTEST, "squash-policy-eu-ai-act.json"], answer_cli=["squash attest --policy eu-ai-act --sign", "squash trust-package"], days_to_close=3, ), ExamQuestion( q_id="EU-035", article="Art. 16(e)", weight=1, category="conformity", question="Has the AI system been registered in the EU AI database as required?", answer_sources=[_ATTEST], answer_cli=["squash publish --registry eu"], days_to_close=5, ), ExamQuestion( q_id="EU-036", article="Art. 12(2)", weight=1, category="record-keeping", question="Are logs collected with sufficient granularity to enable reconstruction of events post-incident?", answer_sources=[_ATTEST, _DRIFT], answer_cli=["squash watch", "squash drift-check --baseline ./model"], days_to_close=2, ), ExamQuestion( q_id="EU-037", article="Annex IV §7", weight=1, category="technical-documentation", question="Are software version control and change management records maintained?", answer_sources=[_BOM, _ATTEST], answer_cli=["squash attest --sign", "squash diff v1.json v2.json"], days_to_close=1, ), ExamQuestion( q_id="EU-038", article="Art. 62(1)", weight=1, category="post-market", question="Is there a post-market monitoring plan with defined KPIs for performance and safety?", answer_sources=[_DRIFT, _ATTEST], answer_cli=["squash monitor", "squash drift-check"], days_to_close=3, ), ] def _nist_rmf_questions() -> list[ExamQuestion]: """30 examiner questions for NIST AI Risk Management Framework 1.0.""" return [ # ── GOVERN function ─────────────────────────────────────────────── ExamQuestion( q_id="NIST-G01", article="GOVERN 1.1", weight=3, category="govern", question="Are policies, processes, and procedures in place that enable organisational risk management decisions for AI?", answer_sources=[_ATTEST, _SQUASH_CFG, "squash-policy-enterprise-strict.json"], answer_cli=["squash init", "squash attest --policy enterprise-strict"], days_to_close=3, ), ExamQuestion( q_id="NIST-G02", article="GOVERN 1.2", weight=2, category="govern", question="Are accountability mechanisms established for AI risk decisions across the organisation?", answer_sources=[_ATTEST, _ANNEX_IV], answer_cli=["squash attest", "squash annex-iv generate"], days_to_close=3, ), ExamQuestion( q_id="NIST-G03", article="GOVERN 1.3", weight=2, category="govern", question="Are organisational roles and responsibilities defined for AI risk management?", answer_sources=[_ANNEX_IV, _MODEL_CARD], answer_cli=["squash annex-iv generate", "squash model-card"], days_to_close=2, ), ExamQuestion( q_id="NIST-G04", article="GOVERN 1.4", weight=2, category="govern", question="Are teams working on AI systems trained on risk and responsible AI policies?", answer_sources=[_SQUASH_CFG, _ATTEST], answer_cli=["squash init", "squash install-hook"], days_to_close=5, ), ExamQuestion( q_id="NIST-G05", article="GOVERN 1.5", weight=2, category="govern", question="Are processes in place for monitoring and reporting on AI risk management effectiveness?", answer_sources=[_DRIFT, _ATTEST, _SCAN], answer_cli=["squash drift-check", "squash watch"], days_to_close=3, ), ExamQuestion( q_id="NIST-G06", article="GOVERN 1.6", weight=1, category="govern", question="Are policies documented for handling AI system failures and incidents?", answer_sources=[_INCIDENT, _SQUASH_CFG], answer_cli=["squash freeze --help", "squash incident"], days_to_close=5, ), ExamQuestion( q_id="NIST-G07", article="GOVERN 4.1", weight=2, category="govern", question="Are third-party AI components and suppliers assessed for risk?", answer_sources=[_BOM, _VEX, _SCAN], answer_cli=["squash attest", "squash vex update", "squash scan ./model"], days_to_close=3, ), ExamQuestion( q_id="NIST-G08", article="GOVERN 6.1", weight=2, category="govern", question="Are policies established for managing concentration risk across AI vendors?", answer_sources=[_BOM, _ATTEST], answer_cli=["squash vendor-registry", "squash attest"], days_to_close=5, ), # ── MAP function ────────────────────────────────────────────────── ExamQuestion( q_id="NIST-M01", article="MAP 1.1", weight=3, category="map", question="Is context established for the AI system including purpose, scope, and stakeholders?", answer_sources=[_ANNEX_IV, _MODEL_CARD], answer_cli=["squash annex-iv generate", "squash model-card --format all"], days_to_close=2, ), ExamQuestion( q_id="NIST-M02", article="MAP 1.5", weight=2, category="map", question="Have organisational risk tolerances been established and documented?", answer_sources=["squash-policy-enterprise-strict.json", _ANNEX_IV], answer_cli=["squash attest --policy enterprise-strict", "squash risk-assess"], days_to_close=3, ), ExamQuestion( q_id="NIST-M03", article="MAP 1.6", weight=2, category="map", question="Have AI system benefits and potential harms been identified and documented?", answer_sources=[_ANNEX_IV, "squash-policy-eu-ai-act.json"], answer_cli=["squash risk-assess", "squash annex-iv generate"], days_to_close=2, ), ExamQuestion( q_id="NIST-M04", article="MAP 2.1", weight=2, category="map", question="Is the scientific basis for the AI model well-understood and documented?", answer_sources=[_ANNEX_IV, _MODEL_CARD], answer_cli=["squash annex-iv generate", "squash model-card"], days_to_close=2, ), ExamQuestion( q_id="NIST-M05", article="MAP 5.1", weight=1, category="map", question="Have practices identified in the NIST AI RMF Playbook been reviewed?", answer_sources=[_NIST_RMF, _ATTEST], answer_cli=["squash attest --policy nist-ai-rmf"], days_to_close=5, ), # ── MEASURE function ────────────────────────────────────────────── ExamQuestion( q_id="NIST-ME01", article="MEASURE 1.1", weight=3, category="measure", question="Are AI risk metrics established and available for testing, monitoring, and evaluation?", answer_sources=[_SCAN, _ATTEST, "squash-policy-enterprise-strict.json"], answer_cli=["squash attest --policy enterprise-strict", "squash scan ./model"], days_to_close=2, ), ExamQuestion( q_id="NIST-ME02", article="MEASURE 2.1", weight=2, category="measure", question="Are test sets and evaluation procedures in place to assess AI system performance?", answer_sources=[_ATTEST, _SCAN, _ANNEX_IV], answer_cli=["squash attest", "squash annex-iv generate"], days_to_close=3, ), ExamQuestion( q_id="NIST-ME03", article="MEASURE 2.5", weight=2, category="measure", question="Are bias and fairness metrics evaluated and documented?", answer_sources=[_BIAS, _ANNEX_IV], answer_cli=["squash bias-audit --model ./model"], days_to_close=5, ), ExamQuestion( q_id="NIST-ME04", article="MEASURE 2.6", weight=2, category="measure", question="Are AI system outputs evaluated for accuracy and consistency over time?", answer_sources=[_DRIFT, _ATTEST], answer_cli=["squash drift-check", "squash monitor"], days_to_close=3, ), ExamQuestion( q_id="NIST-ME05", article="MEASURE 2.8", weight=2, category="measure", question="Are cybersecurity risks to the AI system identified and evaluated?", answer_sources=[_SCAN, _VEX], answer_cli=["squash scan ./model --exit-2-on-unsafe", "squash vex update"], days_to_close=2, ), ExamQuestion( q_id="NIST-ME06", article="MEASURE 2.13", weight=1, category="measure", question="Are effects of AI on people and communities evaluated and documented?", answer_sources=[_BIAS, _ANNEX_IV], answer_cli=["squash bias-audit --model ./model", "squash risk-assess --include-societal"], days_to_close=5, ), ExamQuestion( q_id="NIST-ME07", article="MEASURE 4.1", weight=2, category="measure", question="Is risk-measurement data collected to enable feedback on risk management?", answer_sources=[_DRIFT, _SCAN, _ATTEST], answer_cli=["squash drift-check", "squash watch"], days_to_close=2, ), # ── MANAGE function ─────────────────────────────────────────────── ExamQuestion( q_id="NIST-MA01", article="MANAGE 1.1", weight=3, category="manage", question="Are identified AI risks prioritised and managed using established plans?", answer_sources=["squash-policy-enterprise-strict.json", _ATTEST], answer_cli=["squash attest --policy enterprise-strict", "squash remediate"], days_to_close=2, ), ExamQuestion( q_id="NIST-MA02", article="MANAGE 1.3", weight=2, category="manage", question="Are responses to AI risks established, documented, and operationalised?", answer_sources=[_INCIDENT, _ATTEST], answer_cli=["squash incident", "squash freeze --help"], days_to_close=3, ), ExamQuestion( q_id="NIST-MA03", article="MANAGE 2.2", weight=2, category="manage", question="Are mechanisms in place to monitor and evaluate the effectiveness of risk responses?", answer_sources=[_DRIFT, _ATTEST], answer_cli=["squash watch", "squash drift-check"], days_to_close=2, ), ExamQuestion( q_id="NIST-MA04", article="MANAGE 2.4", weight=2, category="manage", question="Are negative impacts and residual risks documented along with treatment decisions?", answer_sources=[_ANNEX_IV, _ATTEST], answer_cli=["squash risk-assess", "squash annex-iv generate"], days_to_close=2, ), ExamQuestion( q_id="NIST-MA05", article="MANAGE 3.1", weight=2, category="manage", question="Are AI risk management activities tracked and reported?", answer_sources=[_ATTEST, _DRIFT], answer_cli=["squash board-report", "squash annual-review"], days_to_close=2, ), ExamQuestion( q_id="NIST-MA06", article="MANAGE 3.2", weight=1, category="manage", question="Are processes for risk management improvement and learning documented?", answer_sources=[_INCIDENT, _ATTEST], answer_cli=["squash remediate", "squash annual-review"], days_to_close=3, ), ExamQuestion( q_id="NIST-MA07", article="MANAGE 4.1", weight=2, category="manage", question="Are AI systems decommissioned or retired based on risk evaluations?", answer_sources=[_INCIDENT, _ANNEX_IV], answer_cli=["squash freeze", "squash incident"], days_to_close=5, ), ExamQuestion( q_id="NIST-MA08", article="MANAGE 4.2", weight=1, category="manage", question="Are lessons learned from AI incidents documented and incorporated into risk management?", answer_sources=[_INCIDENT, _DRIFT], answer_cli=["squash incident", "squash annual-review"], days_to_close=3, ), ExamQuestion( q_id="NIST-MA09", article="MANAGE 2.1", weight=2, category="manage", question="Are treatment options evaluated and selected based on their effectiveness and feasibility?", answer_sources=[_ATTEST, "squash-policy-enterprise-strict.json"], answer_cli=["squash remediate", "squash attest --policy enterprise-strict"], days_to_close=2, ), ExamQuestion( q_id="NIST-MA10", article="GOVERN 5.1", weight=1, category="manage", question="Are AI risk management processes periodically reviewed and updated?", answer_sources=[_ATTEST, _DRIFT], answer_cli=["squash annual-review", "squash drift-check"], days_to_close=3, ), ] def _sec_questions() -> list[ExamQuestion]: """22 examiner questions for SEC AI disclosure obligations (2026).""" return [ ExamQuestion( q_id="SEC-001", article="AI Disclosure §17", weight=3, category="disclosure", question="Are material AI risks disclosed in investor filings with specific, not generic, risk language?", answer_sources=[_ATTEST, _MODEL_CARD, "squash-policy-eu-ai-act.json"], answer_cli=["squash attest", "squash model-card --format all"], days_to_close=5, ), ExamQuestion( q_id="SEC-002", article="OMB M-26-04 §3", weight=3, category="disclosure", question="Has a model card and acceptable-use policy been prepared for each AI system used in regulated activities?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format all --validate", "squash annex-iv generate"], days_to_close=3, ), ExamQuestion( q_id="SEC-003", article="Investment Adviser Act §204", weight=3, category="compliance", question="Are AI-generated investment recommendations subject to human review with documented oversight controls?", answer_sources=[_ATTEST, _ANNEX_IV], answer_cli=["squash attest --policy eu-ai-act"], days_to_close=5, ), ExamQuestion( q_id="SEC-004", article="AI Claims §11", weight=2, category="disclosure", question="Has the organisation verified that public AI capability claims are substantiated by attestation evidence?", answer_sources=[_ATTEST, _MODEL_CARD], answer_cli=["squash attest --sign", "squash model-card --validate"], days_to_close=3, ), ExamQuestion( q_id="SEC-005", article="Cybersecurity Rule §275", weight=2, category="cybersecurity", question="Are AI model security assessments documented and included in cybersecurity disclosures?", answer_sources=[_SCAN, _VEX, _ATTEST], answer_cli=["squash scan ./model", "squash vex update"], days_to_close=3, ), ExamQuestion( q_id="SEC-006", article="AI Exam Priority §1", weight=2, category="operational", question="Are AI systems that influence investment decisions subject to ongoing performance monitoring?", answer_sources=[_DRIFT, _ATTEST], answer_cli=["squash drift-check", "squash monitor"], days_to_close=3, ), ExamQuestion( q_id="SEC-007", article="Data Governance §15", weight=2, category="data-governance", question="Are training data sources for AI models used in investment activities documented?", answer_sources=[_LINEAGE, _BOM], answer_cli=["squash data-lineage --model ./model"], days_to_close=3, ), ExamQuestion( q_id="SEC-008", article="AI Bias §11(d)", weight=2, category="fairness", question="Have AI models been tested for discriminatory outcomes in investment or advisory functions?", answer_sources=[_BIAS, _ATTEST], answer_cli=["squash bias-audit --model ./model"], days_to_close=5, ), ExamQuestion( q_id="SEC-009", article="Operational Risk §17(a)", weight=2, category="operational", question="Are AI model failures and incidents reported to compliance with appropriate escalation procedures?", answer_sources=[_INCIDENT, _ATTEST], answer_cli=["squash incident", "squash freeze"], days_to_close=5, ), ExamQuestion( q_id="SEC-010", article="Vendor Risk §15(c)", weight=2, category="third-party", question="Are third-party AI model providers subject to due diligence including attestation review?", answer_sources=[_BOM, _VEX, _SCAN], answer_cli=["squash attest", "squash scan ./model", "squash vex update"], days_to_close=3, ), ExamQuestion( q_id="SEC-011", article="Model Risk §SR 11-7", weight=2, category="model-risk", question="Is there a model inventory and validation programme covering AI models used in financial decisions?", answer_sources=[_ATTEST, _BOM, _DRIFT], answer_cli=["squash registry", "squash drift-check"], days_to_close=5, ), ExamQuestion( q_id="SEC-012", article="AI Claims §7(b)", weight=2, category="disclosure", question="Are limitations and failure modes of AI systems disclosed to clients?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format all"], days_to_close=2, ), ExamQuestion( q_id="SEC-013", article="Record Keeping §204(a)", weight=1, category="record-keeping", question="Are records of AI usage in investment decisions retained for the mandated period?", answer_sources=[_ATTEST, _DRIFT], answer_cli=["squash attest --sign", "squash annual-review"], days_to_close=2, ), ExamQuestion( q_id="SEC-014", article="Conflicts §206(3)", weight=2, category="compliance", question="Have conflicts of interest arising from AI use in investment decisions been identified and disclosed?", answer_sources=[_ANNEX_IV, _ATTEST], answer_cli=["squash risk-assess", "squash annex-iv generate"], days_to_close=5, ), ExamQuestion( q_id="SEC-015", article="AI Exam Priority §3", weight=1, category="operational", question="Do compliance policies explicitly address AI use cases within investment operations?", answer_sources=[_SQUASH_CFG, "squash-policy-enterprise-strict.json"], answer_cli=["squash init", "squash attest --policy enterprise-strict"], days_to_close=3, ), ExamQuestion( q_id="SEC-016", article="OMB M-26-04 §4", weight=2, category="disclosure", question="Are AI evaluation artefacts (test results, benchmarks, red-team reports) retained and available for examination?", answer_sources=[_ATTEST, _SCAN, _ANNEX_IV], answer_cli=["squash attest --sign", "squash annex-iv generate"], days_to_close=2, ), ExamQuestion( q_id="SEC-017", article="AI Wash §17", weight=2, category="disclosure", question="Are AI capability claims in marketing materials verified against attestation evidence?", answer_sources=[_ATTEST, _MODEL_CARD], answer_cli=["squash model-card --validate", "squash attest --sign"], days_to_close=3, ), ExamQuestion( q_id="SEC-018", article="BC/DR §17a-4", weight=1, category="operational", question="Is there a business continuity plan for AI system failures affecting investment operations?", answer_sources=[_INCIDENT, _SQUASH_CFG], answer_cli=["squash freeze --help"], days_to_close=10, ), ExamQuestion( q_id="SEC-019", article="Training §204(a)", weight=1, category="operational", question="Are staff using AI in investment activities trained on risks, limits, and compliance obligations?", answer_sources=[_SQUASH_CFG, _ATTEST], answer_cli=["squash init"], days_to_close=10, ), ExamQuestion( q_id="SEC-020", article="Audit Trail §17a-4", weight=2, category="record-keeping", question="Is there an immutable audit trail for AI model versions and decisions?", answer_sources=[_ATTEST, _BOM], answer_cli=["squash attest --sign", "squash diff"], days_to_close=2, ), ExamQuestion( q_id="SEC-021", article="Concentration §206", weight=1, category="operational", question="Is concentration risk from reliance on a single AI vendor assessed and managed?", answer_sources=[_BOM, _ATTEST], answer_cli=["squash attest"], days_to_close=5, ), ExamQuestion( q_id="SEC-022", article="Change Mgmt §17(a)", weight=1, category="operational", question="Are changes to AI models subject to a validation and approval process before deployment?", answer_sources=[_ATTEST, _DRIFT], answer_cli=["squash diff", "squash attest", "squash registry-gate --backend mlflow"], days_to_close=3, ), ] def _fda_questions() -> list[ExamQuestion]: """20 examiner questions for FDA AI/ML Software as a Medical Device (SaMD).""" return [ ExamQuestion( q_id="FDA-001", article="SaMD Rule §880", weight=3, category="risk-classification", question="Has the device software been classified under the correct SaMD risk category (Class I/II/III)?", answer_sources=[_ANNEX_IV, _ATTEST], answer_cli=["squash annex-iv generate", "squash risk-assess"], days_to_close=5, ), ExamQuestion( q_id="FDA-002", article="510(k) §807", weight=3, category="clearance", question="Has the AI/ML system received appropriate FDA clearance or approval, or is it exempt?", answer_sources=[_ATTEST, _ANNEX_IV], answer_cli=["squash attest --sign", "squash trust-package"], days_to_close=30, ), ExamQuestion( q_id="FDA-003", article="SaMD Action Plan §3.1", weight=3, category="clinical-validation", question="Has analytical and clinical validation been conducted and documented for the AI/ML algorithm?", answer_sources=[_ANNEX_IV, _ATTEST, _BIAS], answer_cli=["squash annex-iv generate", "squash bias-audit --model ./model"], days_to_close=10, ), ExamQuestion( q_id="FDA-004", article="AI Action Plan §4", weight=2, category="post-market", question="Is there a predetermined change control plan (PCCP) for planned AI model modifications?", answer_sources=[_DRIFT, _ATTEST, _ANNEX_IV], answer_cli=["squash drift-check", "squash watch"], days_to_close=5, ), ExamQuestion( q_id="FDA-005", article="21 CFR Part 820", weight=3, category="quality-system", question="Is there a Quality Management System (QMS) covering the AI/ML device development lifecycle?", answer_sources=[_ATTEST, _SQUASH_CFG, _BOM], answer_cli=["squash init", "squash attest --sign"], days_to_close=10, ), ExamQuestion( q_id="FDA-006", article="SaMD §4.3", weight=2, category="clinical-validation", question="Are intended use, intended patient population, and clinical context documented?", answer_sources=[_ANNEX_IV, _MODEL_CARD], answer_cli=["squash annex-iv generate", "squash model-card --format all"], days_to_close=3, ), ExamQuestion( q_id="FDA-007", article="AI Action Plan §2", weight=2, category="transparency", question="Is there a labelling strategy that informs users of the AI/ML device's outputs and limitations?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format all"], days_to_close=3, ), ExamQuestion( q_id="FDA-008", article="Post-Market §21 CFR 803", weight=2, category="post-market", question="Are adverse event reporting procedures in place for AI/ML device malfunctions?", answer_sources=[_INCIDENT, _ATTEST], answer_cli=["squash incident", "squash freeze"], days_to_close=5, ), ExamQuestion( q_id="FDA-009", article="SaMD §4.4", weight=2, category="data-governance", question="Are training and testing datasets documented with source, size, demographics, and preprocessing?", answer_sources=[_LINEAGE, _ANNEX_IV, _BIAS], answer_cli=["squash data-lineage --model ./model", "squash annex-iv generate"], days_to_close=5, ), ExamQuestion( q_id="FDA-010", article="Fairness §5.1", weight=2, category="fairness", question="Has the AI/ML model been evaluated for performance disparities across demographic subgroups?", answer_sources=[_BIAS, _ANNEX_IV], answer_cli=["squash bias-audit --model ./model"], days_to_close=5, ), ExamQuestion( q_id="FDA-011", article="Cybersecurity §524B", weight=2, category="cybersecurity", question="Has a cybersecurity risk assessment been conducted and documented for the AI/ML device?", answer_sources=[_SCAN, _VEX], answer_cli=["squash scan ./model --exit-2-on-unsafe", "squash vex update"], days_to_close=5, ), ExamQuestion( q_id="FDA-012", article="SaMD §3.2", weight=2, category="quality-system", question="Is there a software version control and configuration management process?", answer_sources=[_BOM, _ATTEST], answer_cli=["squash attest --sign", "squash diff v1.json v2.json"], days_to_close=2, ), ExamQuestion( q_id="FDA-013", article="AI Action Plan §3.2", weight=2, category="post-market", question="Is real-world performance data collected and analysed for ongoing monitoring?", answer_sources=[_DRIFT, _ATTEST], answer_cli=["squash drift-check", "squash monitor"], days_to_close=3, ), ExamQuestion( q_id="FDA-014", article="SaMD §5.1", weight=1, category="risk-classification", question="Are failure mode analyses (FMEA/FMECA) documented for the AI/ML algorithm?", answer_sources=[_ANNEX_IV, _SCAN], answer_cli=["squash risk-assess", "squash scan ./model"], days_to_close=10, ), ExamQuestion( q_id="FDA-015", article="21 CFR Part 11", weight=1, category="record-keeping", question="Do electronic records meet 21 CFR Part 11 requirements for integrity and audit trail?", answer_sources=[_ATTEST, _BOM], answer_cli=["squash attest --sign"], days_to_close=5, ), ExamQuestion( q_id="FDA-016", article="Labelling §801", weight=1, category="transparency", question="Does device labelling include sufficient information for safe and effective use?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format all"], days_to_close=3, ), ExamQuestion( q_id="FDA-017", article="Human Factors §11937", weight=2, category="usability", question="Has user interface validation been conducted to ensure safe human-device interaction?", answer_sources=[_ANNEX_IV], answer_cli=["squash annex-iv generate"], days_to_close=10, ), ExamQuestion( q_id="FDA-018", article="SaMD §6", weight=1, category="supply-chain", question="Are software component supply-chain risks documented and managed?", answer_sources=[_BOM, _VEX], answer_cli=["squash attest", "squash vex update"], days_to_close=3, ), ExamQuestion( q_id="FDA-019", article="AI Action Plan §5", weight=1, category="transparency", question="Is transparency information about the AI model's training and capabilities publicly available?", answer_sources=[_MODEL_CARD, _ANNEX_IV], answer_cli=["squash model-card --format all --push-to-hub user/model"], days_to_close=3, ), ExamQuestion( q_id="FDA-020", article="AI Action Plan §1", weight=2, category="post-market", question="Has the organisation engaged with FDA's Digital Health Center of Excellence for guidance?", answer_sources=[_ATTEST, _ANNEX_IV], answer_cli=["squash attest --sign"], days_to_close=30, ), ] __all__ = [ "ExamQuestion", "ExamAnswer", "ReadinessReport", "AuditSimulator", "SUPPORTED_REGULATORS", ] SUPPORTED_REGULATORS = AuditSimulator.SUPPORTED_REGULATORS