← All specifications

CMS AI Guidance (TRA) — ponens Policy Pack

This pack maps the CMS Technical Reference Architecture AI Guidance onto computable ponens policies. Where the NIST AI RMF is the generic risk lifecycle, CMS is the operational enforcement layer — concrete business rules and operational-security practices for using AI responsibly with sensitive federal healthcare data.

Source: CMS TRA, Artificial Intelligence Guidance (business rules BR-AI-1..6; references OMB M-25-21/M-25-22, NIST AI RMF, HHS AI policy). https://www.cms.gov/tra/Foundation/FD_0080_Foundation_AI_Guidance.htm

Why this maps onto ponens

CMS publishes six enforceable business rules plus operational practices that map directly to policies over an AI system’s operation record. It is notably self-describing for ponens: CMS mandates tracking “traces, EVALs, prompt management/versioning, and key metrics” — exactly the governance-semantic telemetry ponens evaluates, so this pack is close to a literal implementation of the guidance.

It is distinct from the lifecycle frameworks by being operational and enforceable: high-impact use-case tiering (the OMB M-25-21 definition), PHI/sensitive-data gating, data residency / foreign-AI on-prem only, AI supply-chain provenance, zero-trust for AI, and records retention.

CMS AI Guidanceponens
The AI system’s operation recordthe trace
A business rule (BR-AI-n) / practicea policy (temporal formula)
Mandatory rule vs recommended practiceerror (Red) / warning (Amber)

Trace model

Reuses existing vocabulary — Plan, Decision, Deploy, Output, EditFile (no new action types). Predicates include high_impact_ai, high_impact_decision, risk_assessment_done, human_final_decision, documented_oversight, human_oversight, ai_written_policy, human_review, sensitive_data_use, approved_tool, foreign_ai, on_cms_infrastructure, no_internet_egress, external_ai, data_use_agreement, nonprod_data, synthetic_or_deidentified, ai_component, provenance_verified, ai_outbound_capable, data_minimization, network_segmented, production_ai, prompts_versioned, observability_enabled, ai_supported_action, records_retained.

Worked traces: examples/cms/governed.json (13/13 Green) and violating.json (8 Red + 2 Amber). Run ponens trace check <file>.

The pack

errorRed; warningAmber.

Use-Case Governance & Oversight — BR-AI-2/4/5 (conformance)

PolicyFormulaRAG
cms_high_impact_risk_assessedG(high_impact_ai → P(risk_assessment_done))R
cms_high_impact_human_final_decisionG(high_impact_decision → human_final_decision ∧ documented_oversight)R
cms_ai_policy_human_reviewG(ai_written_policy → human_review)A
cms_continuous_human_oversightG((Deploy ∨ Output) → P(human_oversight))R

Data Protection & Residency — BR-AI-1/3 (security)

PolicyFormulaRAG
cms_sensitive_data_compliant_toolG(sensitive_data_use → approved_tool)R
cms_data_residencyG(foreign_ai → on_cms_infrastructure ∧ no_internet_egress)R
cms_external_ai_data_agreementG(external_ai → P(data_use_agreement))A
cms_privacy_preserving_nonprodG(nonprod_data → synthetic_or_deidentified)A

Supply-Chain & Zero-Trust (security)

PolicyFormulaRAG
cms_ai_provenance_verifiedG(ai_component → provenance_verified)R
cms_zero_trust_for_aiG(ai_outbound_capable → data_minimization ∧ network_segmented)R

Observability & Records — BR-AI-6 (auditability)

PolicyFormulaRAG
cms_production_observabilityG((Deploy ∨ Output) → P(observability_enabled))A
cms_prompts_versionedG(production_ai → prompts_versioned)A
cms_records_retentionG(ai_supported_action → records_retained)R

Aggregation

ponens trace check aggregates the pack: any error fail ⇒ Red (an unassessed high-impact use case, AI making a high-impact final decision, sensitive data on a non-compliant tool, foreign AI egressing data, unverified provenance, or unretained AI-supported actions); else any warning fail ⇒ Amber (policy review, external-AI agreement, privacy-preserving data, observability, prompt versioning); else Green.

Notes

CMS AI Guidance sits on top of the NIST AI RMF (which it references) and OMB M-25-21/22 — it is the agency-level operational instantiation. It pairs with the NIST AI RMF (lifecycle), SSDF (secure build), and FIX (runtime) packs: govern the risk, build securely, operate under the agency rules, gate at runtime.