Evidence ledger
Sources and review notes
The handbook was fact-checked against primary and official sources in August 2026. Product features and policies change, so open the linked source before making a compliance or purchasing decision.
The recipes also contain professional judgment: suggested prompts, review routines and workflow choices. Those are presented as practical guidance, not measured product rankings or legal, compliance, medical, or accreditation advice.
Editorial standard
How the handbook is reviewed
- Claims first. Privacy, accreditation, research, and product claims are checked against primary or official sources.
- Judgment labeled. Prompts, review routines, and workflow choices are practical recommendations, not measured product rankings.
- Dates visible. Product and policy guidance carries a review date because features, terms, and standards change.
- Responsibility stays local. Readers are directed back to their accreditor, organization, privacy lead, or other responsible authority. The ACCME guidebook is a map of the Tuesday jobs, not a substitute for current ACCME wording.
Healthcare privacy
- HHS: HIPAA and cloud computing Business associate agreements, risk analysis and safeguards for ePHI.
- HHS: Business associates Required written assurances, permitted uses and safeguards.
- HHS: Guidance on de-identification The formal Safe Harbor and Expert Determination methods.
Accredited education and research
- ACCME: Standards for Integrity and Independence The five independence standards, in effect 1 January 2022.
- ACCME: Accreditation criteria Core criteria and the optional commendation menu.
- ACCME: Guidance on AI Disclosure, human oversight and commercial-bias review, January 2026.
- ACCME: Identify, mitigate and disclose relevant financial relationships Disclosure to learners before the education.
- ACCME: Joint providership Accredited provider owns compliance. Not a legal partnership. Ineligible companies cannot jointly provide.
- ACCME: Enduring material On-demand format, PARS date range up to three years, verified learners, credit as time to complete.
- ACCME: Content validity of enduring materials Review at least every three years. Release, review, and termination dates on the material.
- ACCME: Annual reporting Five-step PARS close. Confirm this year's deadline. State-accredited providers may finish earlier.
- ACCME: Annual Reporting in PARS checklist Official buttons and fields. Our companion is the checking half.
- ACCME: Regularly scheduled series definition One PARS activity per 12-month series. Credits for the whole series. Learners counted per session attended.
- ACCME: Data reporting quick answers Program Summary categories, multi-year enduring counts, activity statuses.
- ACCME: Maintenance of Certification Register CME for MOC/CC in PARS. Current collaborating-board list and program guide. Do not invent recognition statements.
- ACCME: Alert on AI in accredited continuing education Human and clinical oversight for validity, accuracy and bias. April 2026.
- ACCME: 2025 Data Report Size of the accredited system, funding mix, format mix, and outcome-measurement rates. Released 23 June 2026.
- ACCME: Summer 2026 data updates Collaborating boards and PARS notes, including ABPM CCP. The MOC/CC list moves. Confirm it here, then on the MOC page.
- Moore, Green and Gallis: Achieving desired results and improved outcomes JCEHP, 2009. The outcomes levels this field uses: participation through community health. Level 2 is satisfaction. Level 5 is performance.
- McMahon: Accredited CME delivers Workforce, impact, and independence evidence in the Journal of CME, 2025.
- Alliance Almanac: How CME/CPD professionals are implementing AI 2025 survey: 95 percent use AI; time and training are the barriers.
- Alliance: 2025 AI survey dashboard Interactive cut of the 199-respondent poll, including policy and training.
- Alliance: Position on artificial intelligence in CPD Transparency, privacy, human oversight and disclosure principles.
- ICMJE: AI use by authors Disclosure, accountability and source verification.
- NIH: Confidentiality in peer review Prohibition on uploading confidential applications to generative AI.
- NSF: Generative AI and merit review Reviewer restrictions and disclosure guidance for proposers.
Product capabilities and data controls
- OpenAI: Projects in ChatGPT Files, chats, instructions, plan availability and workspace data treatment.
- OpenAI: Enterprise privacy Business data controls and training defaults.
- OpenAI: HIPAA-eligible products and functionality Eligible services and configurations, not a blanket claim for every plan.
- Anthropic: Claude Projects Project knowledge and instructions. Free accounts can create a small number of projects. Paid plans add retrieval for larger knowledge bases.
- Anthropic: Commercial data roles Claude for Work data processing and consumer-product distinction.
- Microsoft: Data, privacy and security for Microsoft Copilot Permission boundaries and foundation-model training treatment. Licenses may still say Microsoft 365 Copilot.
- Microsoft: How Copilot Notebooks works Reusable, source-grounded notebooks in supported Microsoft 365 plans.
Retrieval, graphs, and loops
- Lewis et al.: Retrieval-augmented generation The 2020 paper that named RAG: retrieve passages, then generate from them.
- Edge et al.: From local to global (GraphRAG) Microsoft Research, 2024. In that study, GraphRAG produced more comprehensive answers on corpus-wide questions. It is not a finding about CME files. Preprint.
- Han et al.: RAG versus GraphRAG 2025 systematic comparison. RAG stronger on single-hop detail; graph methods stronger on some multi-hop questions. Extracted graphs were incomplete. Preprint.
- Yao et al.: ReAct 2022. Reason, act with a tool, observe, repeat. A reference for what “loop” means in agent systems.
- Anthropic: Sycophancy in language models Models tend to agree with the user. Asking the same model whether its draft is right is not an independent check.
Accuracy and citation risk
- NIST: Generative AI Profile Confabulation, false citations and other generative-AI risks.
- Scientific Reports: Fabricated and erroneous citations A primary study demonstrating citation errors in earlier model versions.