NIST AI Risk Management Framework Mapping

Maps the ai-agent-eval-harness-healthtech reference implementation against NIST AI RMF 1.0 (AI 100-1, January 2023). The RMF defines four core functions — Govern, Map, Measure, Manage — each with subcategories. This document groups the repository’s practices under those four functions using its own thematic labels rather than citing specific official subcategory identifiers, and assesses which it addresses today and which require additional work for a production deployment.

Read alongside the regulatory posture and the model card.

Purpose

This is not a NIST AI RMF certification. No certification exists for the RMF; it is a voluntary framework. The purpose of this document is to honestly assess which risk management practices the reference implementation demonstrates and where gaps remain for a production deployment. The assessment is against the repository as-shipped; a forked or deployed instance would need its own assessment.

Framework Mapping

GOVERN — Establish and Maintain AI Risk Management Culture

Theme (our grouping)	Current Implementation	Gap Assessment
Legal and regulatory requirements are understood	Regulatory posture documented in the regulatory posture; FDA wellness/CDS boundary, WHO 2024 guidance, MHRA GMLP, EU AI Act articles mapped	Limited to US/EU/UK/Chile frameworks; a production system would need jurisdiction-specific legal review for each deployment region
AI risk management is embedded in organisational governance	Architecture decision records provide traceability; changes to regulatory posture, guardrails, or escalation require a decision record	No formal governance committee or review board; single-author reference implementation
Roles and responsibilities for AI risk are defined	Clear module ownership: guardrails, eval harness, observability (OpenTelemetry spans per the observability decision)	No separation of duties; author is developer, reviewer, and operator
Risk tolerance is documented and communicated	Documented eval gates, CI-enforced with published pass/fail semantics on faithfulness and hallucination and binary per-case gates on refusal and escalation correctness	Thresholds are binary pass/fail; no graduated risk-acceptance framework
AI systems are transparent	Guardrail-decision trace on every response; citation set on every clinical assertion; model card in CHAI format; regulatory posture publicly accessible	Transparency is at the response level; no public-facing model-performance dashboard
Policies and procedures for AI risk are in place	Security disclosure process; secret scanning in CI; no-secrets policy; PII redaction before LLM	Policies are repository-level, not enterprise-grade; no formal incident-response playbook
Stakeholder engagement	Design informed by published regulatory guidance (FDA, WHO, MHRA); no external stakeholders formally consulted (single-author project)	No patient advisory board, clinical advisory board, or external ethics review

MAP — Understand and Contextualise AI Risks

Theme (our grouping)	Current Implementation	Gap Assessment
Intended purposes and use cases are defined	Model card “Uses and Directions” section; regulatory posture “What the agent does NOT do” list; out-of-scope enumeration enforced by eval harness	Defined for the reference implementation; a deployed product would need context-specific use-case scoping
Interrelated AI risks are identified	Near-miss off-corpus limitation documented in the model card; sub-acute escalation gap acknowledged; generative-model probabilistic behaviour documented	Systematic cross-risk interaction analysis (e.g., how locale bias compounds with near-miss retrieval) is not performed
Constraints and limitations are understood	Honest limitations documented: 45-card single-domain KB, 315-case eval corpus, US-English vocabulary bias, negation-blind escalation, in-memory durability for human-in-the-loop review	Limitations are documented; no formal risk-register with severity scoring
Impact to individuals and groups is assessed	No demographic input features used by the agent; locale-parity scoring addresses cross-locale fairness	No demographic impact assessment beyond locale; no assessment of impact on populations with low health literacy or limited internet access
AI system components are documented	Architecture decision records; the enterprise agent stack; the six-node graph documented in the orchestration decision; model card AI System Facts	Documentation is thorough for the reference implementation; a production system would need operational runbooks
Data provenance is tracked	The data statement and data card document full provenance for the eval corpus and KB cards; source licensing per card	Provenance tracking covers only the shipped synthetic data; no data-lineage tracking for runtime inputs
Third-party risks are identified	The LLM client Protocol abstracts provider dependencies (see the LLM vendor abstraction decision); providers listed in the model card 3rd Party Information	No formal third-party risk assessment; no BAA or contractual review with LLM providers
AI risks are assessed at each lifecycle phase	Eval harness gates every code change; nightly Promptfoo red-team; eval thresholds enforced in CI	Risk assessment is continuous via CI but limited to the eval dimensions scored; no broader organisational risk review at lifecycle gates
Failure modes and cascading impacts are documented	Guardrail failure modes: scope bypass, escalation miss, citation fabrication; each has a test in the eval harness	No formal failure-mode-and-effects analysis (FMEA); cascading impacts across system boundaries not assessed
Stakeholder feedback is incorporated	No external stakeholder feedback loop; design informed by published guidance and author field experience	A production system would need structured feedback channels from patients, clinicians, and compliance officers

MEASURE — Assess and Track AI Risks

Theme (our grouping)	Current Implementation	Gap Assessment
Appropriate metrics are selected	Eight scorer dimensions: citation correctness, citation coverage, refusal correctness, escalation correctness and retrieval recall (deterministic), plus groundedness, faithfulness and hallucination (judge-backed); locale-stratified. Cost and latency are enforced as budget gates rather than scored dimensions. A ninth dimension, escalation / self-harm, is scored by the judge-calibration gate against committed human labels on a separate 81-case corpus	Metrics cover safety and quality; no fairness-specific metrics beyond locale parity; no environmental impact metrics
AI system performance is evaluated	Deterministic CI gate (a key-free deterministic stub client, 315 cases); nightly live-model run; Promptfoo red-team	Deterministic gate is reproducible; live-model metrics are not frozen in the model card (reported in the eval reports)
Evaluation data is representative	315 cases across 3 locales (en, es-419, pt-BR); 5 condition clusters; golden + adversarial + no-match categories	Small sample; no demographic stratification (no demographic data collected); US-English bias acknowledged
Metrics are documented and communicated	Eval reports published per run; model card Key Metrics	Reports are generated per-run; no longitudinal tracking dashboard
Risk thresholds are defined	Hard CI-enforced gates with published pass/fail semantics on faithfulness and hallucination; binary per-case gates on refusal and escalation correctness	Thresholds are crisp but not risk-adjusted; no tiered response framework (e.g., amber vs red)
Monitoring and feedback mechanisms exist	CI eval on every code change; nightly red-team; OpenTelemetry spans on every node; Langfuse Cloud and Phoenix sinks (see the observability decision)	Monitoring covers the reference implementation; no production alerting, SLA monitoring, or degradation-detection pipeline
Bias and fairness are evaluated	Locale parity enforced: identical thresholds for en, es-419, pt-BR; refusal and escalation correctness uniform across locales	No demographic subgroup evaluation (agent takes no demographic input); locale bias limited to vocabulary, not outcome equity
Measurement results are used for improvement	Red-team findings folded into adversarial seed bank; eval regression fails the CI eval check (a signal, not a hard merge gate - no enforced branch protection); release notes track safety-relevant changes	Improvement loop is within-repository; no external audit findings or post-market surveillance data feed

MANAGE — Prioritise and Act on AI Risks

Theme (our grouping)	Current Implementation	Gap Assessment
Risk treatment decisions are documented	Architecture decision records document design decisions affecting risk (the guardrails decision, the observability decision, the eval harness decision, the streaming execution graph decision)	Decision records record design intent; no formal risk-register with treatment plans and residual risk acceptance
AI systems are designed for safe failure	Guardrails fire before LLM (scope classifier, refusal templates, escalation router); citation enforcement refuses on no-match; streaming error events for post-first-byte failures	Near-miss off-corpus handling is a known gap; sub-acute escalation is left to the model
AI risks are mitigated	Deterministic guardrails, eval harness, PII redaction, an OpenTelemetry audit trail, content-negotiated streaming with error events	Mitigations are reference-implementation-grade; production would need additional layers (model supply-chain verification, output filtering at scale)
Incident response plans exist	Security disclosure process; secret scanning; known limitations documented in the model card	No formal incident-response playbook; no on-call rotation; no severity classification scheme
AI system monitoring is ongoing	CI eval on every code change; nightly Promptfoo red-team; OpenTelemetry spans on every turn; cost/latency gates	Monitoring is repository-level; no production alerting, anomaly detection, or automated rollback
AI risks are communicated to stakeholders	The model card, regulatory posture, data statement, and the governance docs in this section are public	Communication is passive (published documents); no active stakeholder notification process for risk changes
Policies and procedures are maintained	Decision-record-based change control; the regulatory posture change-control section requires a decision record for scope changes; release notes track changes	Policies are repository-level; no enterprise policy-management system; no annual policy review cycle

Current State

The reference implementation demonstrates NIST AI RMF practices in the following areas:

GOVERN: Decision-record-based traceability, documented regulatory posture, transparent guardrail decisions, public model card and regulatory posture
MAP: Defined use cases and out-of-scope boundaries, documented data provenance, identified failure modes (near-miss off-corpus, sub-acute escalation), honest limitation statements
MEASURE: Eight-dimension eval harness with deterministic CI gate, locale-stratified scoring, nightly adversarial testing, OpenTelemetry-traced execution
MANAGE: Guardrails-before-LLM architecture, decision-record-documented risk treatment, eval regression that fails the CI eval check (a signal, not a hard merge gate), public disclosure process

The assessment above is honest about what is a reference-implementation demonstration versus a production-grade risk management programme. The four functions are addressed at the depth a reference artefact can reasonably demonstrate: structured documentation, automated measurement, deterministic safety controls, and transparent communication.

What the repository does not have — formal governance committees, risk registers with severity scoring, incident-response playbooks, third-party risk assessments, graduated risk-acceptance frameworks, production monitoring pipelines — is documented explicitly in the Gap Assessment column of each subcategory.

Production Path

A production deployment would need to establish:

Formal governance structure: AI risk committee, defined roles and responsibilities, separation of duties between developers and reviewers, stakeholder advisory boards (patient, clinical, ethics)
Risk register: systematic enumeration of AI risks with severity scoring, likelihood assessment, treatment plans, residual risk acceptance, and risk owner assignment
Expanded measurement: fairness metrics beyond locale parity, environmental impact assessment, demographic subgroup evaluation where applicable, longitudinal performance tracking, automated drift detection
Incident response: formal IR playbook with severity classification, escalation paths, communication templates, post-incident review process, regulatory notification procedures
Third-party governance: vendor risk assessments for LLM providers, BAA where applicable, contractual review for data processing, supply-chain verification for model provenance
Continuous monitoring: production alerting, anomaly detection, automated rollback, degradation early-warning, SLA monitoring, capacity planning
Audit readiness: evidence collection automation, audit log retention (6 years for HIPAA, as applicable), tamper-evident logging, query interface for auditors

The repository’s patterns — eval contracts, decision-record traceability, OpenTelemetry instrumentation, guardrail-first architecture — accelerate building each of these capabilities. They are the foundation, not the finished structure.