SLM Sprint · Governance Briefing
Governing Small, Mid-Size and Large Language Models in Healthcare
Major findings and requirements across task taxonomy, organizational governance, disclosure and risk, and what each means for SLM, MLM and LLM deployment.
SLM
~1–14B · open weights · post-trained by you
MLM
~14–70B · mid-size · on-prem or private cloud
LLM
Frontier · tens to hundreds of billions · vendor API
Darrell A. Hall, MD, MSCIS · Empowered Disease Management Health Services, LLC
October 2026
KEY MESSAGE: Governance, not model quality, is now the gating factor for clinical language models. This briefing gives a size-aware way to decide what to deploy, what to demand from vendors, and what to build in-house.
AUDIENCE FRAME: Written for CMIOs, CIOs, chief AI officers and the technical leads who sit on AI, informatics or P&T committees. It assumes familiarity with EHR integration, model evaluation and HIPAA; no ML math is required.
ILLUSTRATION: Think of the three tiers as formulary classes. An SLM is a compounded preparation: you control every ingredient, so you own the documentation and the liability. An LLM is a branded product: broad indications, but you rely on the manufacturer’s label. An MLM is a licensed generic made in your own facility, with more control than the brand and more breadth than the compound.
TERMS: SLM means roughly 1 to 14 billion parameters, open weights, post-trained in-house. MLM means roughly 14 to 70 billion, self-hosted. LLM means frontier models, usually reached through a vendor API. These cutoffs are working definitions for this briefing; no regulator classifies models by size, and the device boundary depends on intended use.
ARC OF THE TALK: Then (2019 to 2023): one-off validation and licensing-exam benchmarks. Now (2024 to 2026): organizational certification, procurement-driven disclosure, and state laws on patient-facing AI. Next (2027 to 2028): EU high-risk obligations, federal and state preemption fights, and evidence expectations moving into contracts.
FORMAT NOTE: Five slides carry live animations (the policy timeline, the RUAIH consolidation, the evidence shift, drift and change control, and post-training). Let each run one full cycle before speaking to it.
DECISION ASK: By the end, the committee should be able to say which tier fits each use case and what evidence package it will require before go-live.
Finding 1 · The landscape
Governance has split into three layers
LAYER 1
Task taxonomy
What the model does, task by task
MedHELM (Stanford CRFM; Nature Medicine)
LAYER 2
Organizational governance
How the deploying site manages AI
TJC/CHAI RUAIH guidance (2025) and certification (2026)
LAYER 3
Transparency and risk
What to disclose and what to test
CHAI Model Card · HTI-1 DSI · FDA CDS 2026 · NIST AI 600-1
Out of step: MedQA, PubMedQA and the 2023 FDA framing no longer match how hospitals judge models.
Key distinction: only MedHELM evaluates the model. RUAIH certifies the hospital’s process.
Healthcare LM Governance · SLM Sprint · Oct 2026
2
KEY MESSAGE: Three questions, three different owners. What does the model do (the developer)? How does the site govern it (the deploying organization)? What must be disclosed and tested (both parties, under regulation and contract)?
ILLUSTRATION: Use the medication analogy. The task taxonomy is the drug’s list of indications. Organizational governance is the P&T committee and the medication-safety program. Transparency and risk are the package insert and post-market surveillance. An excellent drug in a hospital with no P&T process is still unsafe, and a strong P&T process cannot rescue a drug with no label.
EXAMPLE: An ambient-scribe vendor can show strong note-generation results (layer 1). The hospital still needs a policy on who signs the AI-drafted note and how errors are reported (layer 2), and a model card that says which specialties, languages and visit types were tested (layer 3). Missing any one layer is a gap a surveyor or plaintiff will find.
THEN → NOW: In 2023 most programs evaluated models on licensing-exam style benchmarks such as MedQA and PubMedQA. Frontier models now score near the ceiling on those, so the numbers no longer separate products. The field has moved to task-level evaluation (layer 1) and process-level governance (layer 2), with disclosure (layer 3) in flux.
NEXT (presenter’s read): Layers 2 and 3 converge in procurement, with buyers asking for certification-ready evidence as a contract deliverable rather than a marketing document.
DECISION ASK: Name an accountable owner for each layer before evaluating any product.
Finding 2 · Frameworks in force
Seven instruments now govern clinical AI
Framework Layer Status (Oct 2026) MedHELM (Stanford CRFM) Task taxonomy Active leaderboard; runs locally TJC/CHAI RUAIH guidance Org governance 17 Sep 2025; voluntary TJC RUAIH certification Org governance June 2026; first site 29 Jul CHAI Applied Model Card Disclosure Draft; public registry ONC HTI-1 DSI §170.315(b)(11) Regulatory disclosure HTI-5 proposes removal; verify FDA CDS Software guidance Device boundary 6 Jan 2026; supersedes 2022 NIST AI 600-1 GenAI Profile Risk catalog July 2024; AI RMF in revision
Only MedHELM is a task taxonomy in the ML sense. The rest are schemas a model developer must be able to fill in for any model they post-train.
Healthcare LM Governance · SLM Sprint · Oct 2026
3
KEY MESSAGE: Seven instruments matter today and only one of them evaluates the model. The rest define what the developer must fill in and what the deploying site must run.
WALK THE TABLE: MedHELM is the task benchmark (5 categories, 22 subcategories, 121 tasks, 35 benchmarks). RUAIH guidance (Sep 2025, 7 elements) and RUAIH certification (launched June 2026, 5 areas) govern the deploying organization; Hackensack Meridian was first certified on 29 Jul 2026. The CHAI Applied Model Card is a draft disclosure schema with a public registry. The ONC HTI-1 decision support criterion requires 31 source attributes for predictive tools in certified EHRs. The FDA CDS guidance (6 Jan 2026) sets the device line. NIST AI 600-1 is the generative-AI risk catalog (12 risks across Govern, Map, Measure, Manage).
STATUS CHECK (as of 20 Oct 2026): HTI-5 is still a proposed rule. It went to OMB review on 28 Jul 2026; until a final rule publishes, the HTI-1 requirements stay in force. Present both outcomes as planning scenarios rather than predicting the date.
ILLUSTRATION: Think of building codes. One standard sets the load test for the beam (MedHELM). The others govern the inspector, the permit file, the fire plan and the occupancy rules. Passing the beam test does not get you an occupancy permit.
WHY SEVEN ROWS: The source document listed six frameworks. RUAIH guidance and RUAIH certification are separate instruments with different structures, so they get separate rows here.
DECISION ASK: Decide which of these the organization will treat as mandatory internal policy regardless of what regulators require. The next slide shows why that choice protects you from policy swings.
Policy in motion · 2019 to 2028
Policy is moving in three directions at once
Status as of 20 Oct 2026 · projections are the presenter’s read, not certainties
4
KEY MESSAGE: Policy is moving in different directions at once. US federal health-IT rules are loosening, accreditation and the states are tightening, and the EU has delayed its obligations without dropping them. Plan for the strictest regime you will operate under.
HOW TO PRESENT THE ANIMATION: A cursor sweeps from 2019 to 2028 and the caption box narrates each event as it passes. Solid dots happened; outlined dots to the right of the orange Today line are scheduled dates or projections. Slow down over 2023 to 2026, where most of the movement is. The loop restarts after a short hold.
EVIDENCE LANE (why policy moved): 2019, a widely used care-management algorithm under-identified Black patients because it predicted cost as a proxy for need (Obermeyer et al., Science). 2021, external validation of a proprietary sepsis model found an AUC of 0.63, well below the vendor’s figures (Wong et al., JAMA Internal Medicine). 2024, Texas reached the first state attorney-general settlement over a generative-AI clinical product’s accuracy claims (Pieces Technologies). 2025, MedHELM showed exam scores do not predict task performance. Each failure produced a governance response.
FEDERAL: HTI-1 (Dec 2023) required 31 source attributes for predictive decision support in certified EHRs. Section 1557’s 45 CFR 92.210 has required reasonable efforts against discrimination in patient care decision support tools since May 2025. HTI-5 (proposed Dec 2025) would remove the model-card mandate; it went to OMB review on 28 Jul 2026 and was still not final as of mid-October. EO 14365 (Dec 2025) set up federal challenges to state AI laws.
FDA: 2021 action plan and Good Machine Learning Practice; final PCCP guidance (Dec 2024) lets cleared devices pre-specify updates; revised CDS guidance (Jan 2026) widens the non-device path for clinician-facing tools and stays silent on patient-facing ones.
ACCREDITATION: RUAIH guidance Sep 2025, certification June 2026, first certified system July 2026.
STATES: Utah disclosure (2024); Illinois limits on AI therapy (Aug 2025); California and Texas rules effective Jan 2026; Colorado, Maine, Rhode Island, Tennessee and Vermont restricted AI therapy chatbots in 2026; Colorado’s replacement AI law (SB 26-189) takes effect Jan 2027.
EU: AI Act in force Aug 2024; the Digital Omnibus (in force 27 Jul 2026) moved stand-alone high-risk obligations to 2 Dec 2027 and AI in medical devices to 2 Aug 2028.
PROJECTION (presenter’s read): HTI-5 finalizes in some form; buyers fill the transparency gap through contracts; state patient-facing chatbot laws keep spreading while preemption is litigated.
DECISION ASK: Adopt the strictest applicable requirement as the internal baseline, so swings in federal policy do not force rework.
Finding 3 · Task taxonomy
MedHELM replaces exam scores as headline
5 categories · 22 subcategories · 121 tasks · 35 benchmarks
Clinical Decision Support: diagnosis, treatment planning, risk prediction, knowledge support
Clinical Note Generation: visits, procedures, diagnostic reports, care plans
Patient Communication and Education: resources, care instructions, messaging, accessibility, engagement
Medical Research Assistance: literature and clinical-data analysis
Administration and Workflow: scheduling and coordination
0.47
ICC, LLM-jury vs. clinicians
0.43
ICC, clinician vs. clinician
The jury beat inter-clinician agreement, and exam scores did not predict task performance .
Requirement: lead with MedHELM subsets for target subcategories; keep MedQA/PubMedQA as a knowledge floor.
Healthcare LM Governance · SLM Sprint · Oct 2026
5
KEY MESSAGE: Stop selecting models on exam scores. Select on the tasks you will actually run.
ILLUSTRATION: Board scores predict who passes boards, not who writes a safe discharge summary at 3 a.m. MedHELM tests the discharge summary.
WHAT IT IS: Stanford CRFM’s holistic evaluation, published in Nature Medicine: 5 categories, 22 subcategories, 121 tasks and 35 benchmarks, scored partly by an LLM jury. The jury agreed with clinicians (ICC 0.47) better than clinicians agreed with one another (0.43). That makes scaled evaluation credible, provided you calibrate the jury against your own clinicians. It also tells you clinicians disagree with each other more than we like to admit, so a single physician reviewer is not a gold standard.
EXAMPLE: For a patient-messaging assistant, run the patient-provider messaging and personalized care-instruction tasks on the base model and on the tuned model. Then have clinicians score 50 local outputs by hand and compare their scores with the jury’s. If they diverge, trust neither number until you understand why.
THEN → NOW → NEXT: 2023, MedQA accuracy was the headline and vendors competed on it. 2025, task benchmarks with jury scoring. Projected (presenter’s read): buyers ask for subcategory-level results plus a local replication sample as a standard RFP item.
IMPLEMENTATION POINT: MedHELM can run on local data inside the firewall, so PHI never needs to leave the site. That matters most for SLMs and for behavioral-health and SUD data.
WHERE MedQA STILL BELONGS: As a regression check. If a fine-tune drops general medical knowledge, MedQA will show it quickly and cheaply.
DECISION ASK: Require subcategory-matched results and a local replication sample before any pilot.
Finding 4 · Organizational governance
RUAIH certifies the process, not the model
Size: not split by size, but TJC weighs an applicant’s size and resources.
Requirement: every model must ship evidence a site can use to support certification.
Element-to-area grouping is illustrative · First certified: Hackensack Meridian, 29 Jul 2026
6
KEY MESSAGE: RUAIH certifies the deploying organization’s process, not any model. A vendor cannot be “RUAIH certified”; it can only make your certification easier or harder.
HOW TO PRESENT THE ANIMATION: Each of the seven 2025 guidance elements on the left flows, one at a time, into one of the five 2026 certification areas on the right, which fill in teal as they receive evidence. Say clearly that this grouping is the presenter’s reading of how the elements consolidate, not The Joint Commission’s official crosswalk. Note that monitoring absorbs both quality monitoring and safety-event reporting, and transparency/education absorbs both patient transparency and staff training.
ILLUSTRATION: This works like medication-management accreditation. Surveyors examine your formulary process, storage, error reporting and staff competency; they do not certify that any particular drug is good.
EXAMPLE: Hackensack Meridian Health became the first health system certified, on 29 July 2026. What a surveyor will look for, mapped to the five areas: a governance body with real authority; a data-management policy covering what data models see and keep; documented bias and risk review before deployment; ongoing monitoring with an escalation path and AI events flowing into the blinded safety-reporting system; and staff education plus patient-facing transparency about where AI is used.
SIZE: Certification is not split into small and large tracks, but The Joint Commission has said it takes an applicant’s size and resources into account. For a clinic with one clinical lead, the vendor’s evidence effectively becomes your evidence.
THEN → NOW → NEXT: Before 2025 there was no accreditation standard specific to AI; programs borrowed from IT security and quality frameworks. Now there is guidance plus a voluntary certification. Projection (presenter’s read): payers and large purchasers start asking partners whether they are RUAIH-certified, much as they ask about accreditation today, and some states reference it in AI procurement rules.
DECISION ASK: Map every vendor deliverable to one of the five areas. Anything that maps to none of them is not helping your certification and should not be on the critical path.
Finding 5 · Disclosure and device boundary
Disclosure is moving into procurement
CHAI Applied Model Card
DRAFT · JSON SCHEMA · REGISTRY
Developer · intended use · target population · model type · data types · performance · security · maintenance · risks and out-of-scope uses · bias · ethics
ONC HTI-1 DSI
STATUS TO VERIFY
Source attributes and risk management for predictive DSIs in certified EHRs.
HTI-5 proposes removing both; buyers would then ask for CHAI-style cards.
FDA CDS guidance
6 JAN 2026
4 non-device criteria for clinician-facing tools whose basis can be reviewed.
Silent on patient-facing chatbots and symptom checkers.
Implication: a patient-facing assistant (e.g., PMSG) sits outside the CDS safe harbor at any size. Intended use, not parameter count, sets the device boundary.
Healthcare LM Governance · SLM Sprint · Oct 2026
7
KEY MESSAGE: The legal requirement to publish model information is weakening; the business need to obtain it is not. Disclosure is moving from federal certification into contracts.
WALK THE CARDS: The CHAI Applied Model Card covers intended use, target population, data, performance, maintenance, known risks and out-of-scope uses. An in-house SLM team can fill every field. The HTI-1 decision support criterion requires 31 source attributes in certified EHRs; HTI-5 proposes removing them, citing no public evidence they improved care and low clinician use of the attributes in workflow. The FDA CDS guidance of January 2026 allows a single recommendation under enforcement discretion when only one option is clinically appropriate, moves the time-critical concern from criterion 3 to criterion 4, and remains silent on patient-facing chatbots and on AI specifically.
ILLUSTRATION: A nutrition label few people read at the shelf still matters at purchasing. ONC found clinicians rarely opened source attributes; procurement teams, risk officers and governance committees are the real audience.
WHAT STAYS IN FORCE: Section 1557 (45 CFR 92.210) still requires covered entities to make reasonable efforts to identify and mitigate discrimination risk in patient care decision support tools. Without HTI-1 attributes, you must get those inputs from vendors directly.
EXAMPLE: A patient-facing pain-management assistant (such as the PMSG assistant) that receives “can I take more?” sits outside the CDS safe harbor at any model size, because the user is the patient, not a clinician who can independently review the basis.
DECISION ASK: Write model-card delivery into every AI contract now. The next slide animates why.
Policy change · Who carries the evidence
Under HTI-5, every buyer becomes the tester
HTI-5 is a proposed rule (at OMB since Jul 2026) · Healthcare LM Governance
8
KEY MESSAGE: If HTI-5 is finalized as proposed, one federal test is replaced by many local ones. Every buyer becomes its own certification body, and the five contract asks at the bottom of the slide become the new transparency regime.
HOW TO PRESENT THE ANIMATION: Left panel, today: the vendor is tested once through ONC-authorized certification, and every buyer inherits that single result. One evidence package covers all three buyers. Right panel, HTI-5 as proposed: the same vendor answers each buyer separately. Watch the packets accumulate buyer by buyer, five asks each, and the "runs own test" tag appear under each buyer. The counter reaching 15 is simple arithmetic (3 buyers × 5 asks), not a measured statistic.
ILLUSTRATION: It is the difference between a UL listing on an appliance and every hospital running its own electrical safety test on every appliance it buys.
WHY ONC PROPOSED IT: ASTP/ONC cited no public evidence that the 31 source attributes improved care, and found clinicians rarely opened them in workflow. That is a fair point about the clinician audience. The real audience was always procurement, risk and governance teams, and they still need the information.
WHAT STAYS IN FORCE REGARDLESS: Section 1557 (45 CFR 92.210) still requires covered entities to make reasonable efforts to identify and mitigate discrimination risk in patient care decision support tools. Without the HTI-1 attributes, you must get the inputs for that analysis from the vendor directly.
IMPLICATIONS BY SIZE: Large academic centers can absorb this; they already validate locally. Small hospitals and clinics cannot run subgroup analyses on their volumes, so contract language and third-party evidence (the CHAI registry, RUAIH-certified partners) carry the weight. Vendors will face duplicated questionnaires, and many will publish a standard card to cut that cost, which is the market rebuilding HTI-1 voluntarily.
SLM ANGLE: An in-house SLM team answers its own five asks. The cost is documentation discipline, not negotiation.
PROJECTION (presenter’s read): Standard AI procurement addenda from health-system coalitions, and insurers asking for the same evidence before covering AI-related liability.
DECISION ASK: Adopt a standard five-item AI evidence addendum (model card, subgroup performance, change notification, testing rights, indemnification) for every AI contract, starting now rather than waiting for the final rule.
Finding 6 · NIST AI 600-1 risk catalog
Four of twelve NIST risks dominate
Confabulation
Human-AI configuration
Data privacy
Value chain and components
CBRN information
Dangerous or hateful content
Environmental impact
Harmful bias, homogenization
Information integrity
Information security
Intellectual property
Obscene or abusive content
Confabulation: higher in SLMs on open tasks; contain with scope, retrieval, abstention
Human-AI: fluency invites over-trust at any size; narrow scope eases review
Privacy: LLM PHI leaves the site unless private-hosted; SLM/MLM stay on-prem
Value chain: vendor changes vs. pinned weights and hashed adapters
Healthcare LM Governance · SLM Sprint · Oct 2026
9
KEY MESSAGE: Twelve risks exist; four drive clinical reviews. Spend review time there.
CONFABULATION: Fluent, wrong output. Example: a visit summary that lists a medication the patient never took. Containment: narrow scope, retrieval from the chart so every statement has a source, and trained abstention (“I’m not sure; routing this to your nurse”).
HUMAN-AI CONFIGURATION: Automation bias. Example: clinicians accept AI-drafted patient replies because they read well. Control: put the source next to the draft in the review screen and audit acceptance rates. A near-100% unedited acceptance rate is a warning sign, not a success metric.
DATA PRIVACY: Where PHI goes. A vendor API means PHI leaves the site unless the model is privately hosted under a BAA. SUD records add 42 CFR Part 2, whose updated rule had a compliance date of 16 Feb 2026.
VALUE CHAIN: Components you do not control: the base model’s license, third-party adapters, a vendor quietly switching its underlying model. Control: pinned weights and hashed adapters for anything self-hosted, and change-notification clauses for anything that is not.
ILLUSTRATION: The other eight risks matter, the way every drug interaction matters, but a pharmacist reviewing a new order checks the few that are likely and severe first.
THEN → NOW → NEXT: NIST AI RMF 1.0 (Jan 2023), the generative-AI profile 600-1 (July 2024), and a revision of the framework in progress. Projection (presenter’s read): NIST stays the common vocabulary for US enterprise reviews even as binding federal rules loosen, because it is what auditors and insurers already use.
DECISION ASK: Make a four-row NIST worksheet, with a named control and owner per row, mandatory at intake.
Finding 7 · Care settings (1 of 2)
Size changes who governs and validates
Large AMC Small / critical access Ambulatory / BH Urgent care Emergency dept. Governance AI committee, CMIO/CAIO; RUAIH realistic Quality or P&T committee; leans on vendor One clinical lead plus vendor Set centrally by the chain Inherits hospital committee Who validates Site, on local data Vendor; samples too small Vendor model card only Chain, pooled across sites Site, prospectively Monitoring Continuous drift and subgroup dashboards Periodic audit Chart-sample review Chain QA sampling Safety events via ED quality
Large centers validate locally. Smaller settings inherit vendor evidence, so the model card is the validation they get.
Healthcare LM Governance · SLM Sprint · Oct 2026
10
KEY MESSAGE: “Who validates?” has a different answer in each care setting, and that answer decides which evidence you must buy.
ILLUSTRATION: A large academic center is a hospital pharmacy that can compound and test in-house. A critical-access hospital buys finished product and relies on the label. A clinic has the label and nothing else. Ambulatory/BH in the table means ambulatory, primary care and behavioral health (the teal column).
EXAMPLE: The external validation of a proprietary sepsis model (Wong et al., JAMA Internal Medicine, 2021) is the cautionary tale: it looked strong in the vendor’s data and much weaker at an outside health system. A large center can catch that with a silent trial before go-live. A 25-bed hospital usually cannot, because its volume is too small for stable subgroup estimates.
URGENT CARE AND ED: Urgent-care chains validate centrally, pooling data across sites, which gives them more power than any single clinic. EDs validate prospectively because errors are time-critical, and they fold AI events into existing ED quality review.
IMPLEMENTATION POINT: When the site cannot validate, the model card is the validation. That is why card completeness matters more as you move down-market, and why SLM developers who can fill every field have an advantage there.
PROJECTION (presenter’s read): Pooled or consortium validation services for smaller hospitals, similar to how reference labs serve them today.
DECISION ASK: Classify each deployment site by its validation capacity before choosing products.
Finding 7 · Care settings (2 of 2)
Size changes compute, model fit and law
Large AMC Small / critical access Ambulatory / BH Urgent care Emergency dept. Compute On-prem GPU; 70B-class feasible EHR SaaS; one GPU at most 3–8B SLM on one GPU SaaS Low-latency, on-prem Best-fit model LLM for reasoning; SLM/MLM for volume Vendor SLM or LLM SLM SLM SLM; LLM with clinician review Beyond HIPAA TJC; FDA if device; HTI TJC or CMS CoPs; state AI laws 42 CFR Part 2; state AI and telehealth Corporate policy; state laws FDA time-critical caveat; EMTALA
Where EDMHS sits: PMSG, KATA and the SUD geocoder are patient-facing, run on small infrastructure and touch Part 2 data. That is where an SLM with a strong model card is most credible.
Healthcare LM Governance · SLM Sprint · Oct 2026
11
KEY MESSAGE: Compute and law both push smaller, patient-facing settings toward SLMs.
COMPUTE: A 70B-class model needs multi-GPU infrastructure. A 3 to 8B model quantized to 4-bit runs on a single workstation GPU or on-device. That gap decides whether PHI ever leaves the building.
LAW: Ambulatory and behavioral health carry the densest patient-facing rules. 42 CFR Part 2 covers SUD records. Illinois (Aug 2025) bars AI from making independent therapeutic decisions; Colorado, Maine, Rhode Island, Tennessee and Vermont restricted AI therapy chatbots in 2026. California requires a generative-AI disclaimer and a path to a human for patient communications, and bars AI from implying it is a licensed clinician. Texas requires disclosure of AI use in care. Urgent care inherits corporate policy plus state law; EDs add the EMTALA context and FDA’s time-critical caveat.
DOMINANT TASKS BY SETTING: Large centers use all five MedHELM categories; small hospitals mostly note generation and administration; ambulatory and behavioral health add patient communication and prior authorization; urgent care is notes and discharge instructions; EDs add risk and triage decision support.
EXAMPLE (EDMHS): PMSG, KATA and the SUD geocoder are patient-facing, run on small infrastructure and sit close to Part 2 data. That is the profile where a well-documented SLM is both cheaper to run and easier to defend.
PROJECTION (presenter’s read): More states regulate patient-facing chatbots before federal preemption is settled. Build to the strictest state you serve.
DECISION ASK: Default to SLMs for patient-facing ambulatory and behavioral-health uses unless the task truly needs frontier reasoning.
Model tiers
Control falls as model size rises
SLM
~1–14B · open weights · post-trained by you
Runs on: one GPU; 4-bit on edge
Scope: 1–2 MedHELM categories
Card: every CHAI field fillable
PHI: on-prem or on-device
Fits: ambulatory, BH, urgent care
MLM
~14–70B · open or licensed weights
Runs on: multi-GPU, on-prem or private cloud
Scope: 2–3 categories
Card: full if weights are open
PHI: on-prem if self-hosted
Fits: large systems, high-volume tasks
LLM
Frontier · tens to hundreds of billions
Runs on: vendor API or large cluster
Scope: all 5 categories
Card: limited by vendor disclosure
PHI: leaves site unless private; BAA
Fits: complex reasoning, local validation
No standard size cutoffs; the MLM tier is interpolated between the source’s SLM and LLM definitions
12
KEY MESSAGE: Control falls as size rises, and breadth rises with it. Choose the smallest model that covers the task’s MedHELM subcategories.
ILLUSTRATION: It is the choice between an in-house lab test and a reference lab. In-house is faster, cheaper and fully documented, but runs a narrow menu. The reference lab runs everything, on its own schedule and to its own specifications.
EXAMPLES BY TIER: SLM: patient-message drafting, after-visit summaries, prior-authorization form extraction. MLM: high-volume summarization across many specialties at a large system that self-hosts. LLM: differential-diagnosis support or literature synthesis, always with clinician review.
DEFINITIONS AND CAVEAT: The source document defines SLMs as roughly 1 to 14B, open weights, post-trained by you, and LLMs as frontier models. The MLM tier (about 14 to 70B) is an interpolation for this briefing; 70B-class is what a large academic center can host on-prem. Distillation blurs the lines: a small model trained on a large model’s outputs inherits the teacher’s errors and its license terms.
TREND (presenter’s read): Small models have improved quickly through distillation and preference tuning, so expect tasks to move down a tier each year while frontier models move into multi-step, agentic work, which raises new governance questions of its own.
DECISION ASK: Require a “why not smaller?” justification for any proposal that uses a frontier LLM.
LLM vs. SLM · Dimensions 1–4
Building and evaluating the model
Dimension LLM (frontier / large) SLM (post-trained, ~1–14B) 1 · Architecture and scale Multi-GPU or API; long context; multimodal, native reasoning One GPU or edge; retrieval for context; reasoning by distillation 2 · Training and adaptation Prompting, retrieval, vendor fine-tuning; data opaque SFT then DPO/KTO/IPO via LoRA, QLoRA, DoRA; data documented 3 · Task scope All 5 categories; diagnosis support, research synthesis 1–2 categories; summaries, extraction, messaging 4 · Evaluation MedQA near ceiling; leaderboard; often overconfident MedHELM subsets plus local set; trained abstention
Healthcare LM Governance · SLM Sprint · Oct 2026
13
KEY MESSAGE: How a model is built determines what you can know about it.
1 · ARCHITECTURE AND SCALE: Long context and multimodality favor LLMs, which can read a whole chart or a stack of papers. SLMs need retrieval and chunking to see the same material, which is more engineering but also makes every input traceable.
2 · TRAINING AND ADAPTATION: You rarely retrain a vendor LLM; you prompt it, ground it with retrieval, or use the vendor’s fine-tuning service. With an SLM you can run SFT and then DPO with LoRA adapters on a single GPU and document every training example. Distillation (training a small student on a large teacher’s outputs) is common; the student inherits the teacher’s errors and license terms.
3 · TASK SCOPE: An SLM should own one or two MedHELM categories. Example: a patient-messaging SLM should not be quietly reused for diagnostic support; that is scope creep with no validation behind it.
4 · EVALUATION: Frontier models saturate MedQA, so keep it as a regression check only. For SLMs, measure calibration and abstention explicitly: how often the model says “I don’t know” when it should, and how often it answers when it shouldn’t. Measure subgroup performance on your own demographics, and calibrate the LLM jury against your own clinicians.
ILLUSTRATION: Provenance is a chain of custody for lab specimens. With an SLM you hold the whole chain. With a vendor LLM you receive a sealed result and a certificate.
DECISION ASK: Require documented training-data provenance for any model that touches Part 2 data.
LLM vs. SLM · Dimensions 5–8
Risk, operations and oversight
Dimension LLM (frontier / large) SLM (post-trained, ~1–14B) 5 · Safety (NIST) PHI leaves site unless private; vendor changes uncontrolled PHI on-prem; pinned weights; scope limits confabulation 6 · Deployment Network-bound; per-token pricing vLLM, llama.cpp, TGI; fixed cost; re-validate after quantizing 7 · Governance Card limited by vendor; silent updates; output-only monitoring Every CHAI field; card per version; logit and embedding drift 8 · Regulatory Boundary set by intended use; BAA; Part 2 consent limits Same boundary; local processing eases Part 2 and residency
MLM: SLM-like control if self-hosted with open weights, near-LLM breadth, at the cost of multi-GPU infrastructure and heavier re-validation.
Healthcare LM Governance · SLM Sprint · Oct 2026
14
KEY MESSAGE: Operations is where small and large models diverge most, and where implementation choices pay off.
5 · SAFETY: Narrow scope plus trained abstention contains SLM confabulation. Vendor LLMs confabulate less on open tasks but change outside your control. Human-AI configuration is the same risk at any size: fluency invites over-trust.
6 · DEPLOYMENT AND OPERATIONS: vLLM, llama.cpp or TGI give a predictable fixed cost on owned hardware, versus per-token pricing and network latency for an API. Every compression step (INT8, INT4 quantization, pruning) produces a new model that needs re-validation; treat it like a formulation change.
7 · GOVERNANCE AND TRANSPARENCY: A vendor’s silent model update is a change you did not approve. With pinned SLM weights, a new version means a new model card. Monitoring can go beyond outputs to internal signals such as logit confidence and embedding drift, which give earlier warning.
8 · REGULATORY: The device boundary follows intended use, not size; under the 2026 FDA guidance, patient-facing use falls outside the CDS safe harbor regardless of model size. Local processing simplifies Part 2 consent and state data-residency questions.
MLM: Gets SLM-like control if self-hosted with open weights, at the cost of multi-GPU infrastructure and heavier re-validation per release.
BRIDGE: The next slide animates what these differences look like over a year of operation.
DECISION ASK: Contractually require advance notice of any change to an underlying model.
Implementation · Change control over a year
Change control decides how fast you see a drop
Healthcare LM Governance · SLM Sprint · Oct 2026
15
KEY MESSAGE: Performance is not a launch metric; it is a time series. Change control decides whether a drop is caught in days or in months.
HOW TO PRESENT THE ANIMATION: Both lines start at the same validated level at go-live. The orange line is a vendor-hosted LLM: at month 7 the vendor silently swaps the underlying model and performance falls below the dashed minimum. Output-only review (monthly chart sampling) catches it at month 9, so the shaded band is roughly two months of below-threshold output nobody saw. The teal line is a pinned SLM: a planned release at month 6 passes a revalidation gate before going live (the diamond); gradual input drift starts around month 7 and an embedding-drift alarm fires at month 8, before outputs cross the threshold, and the team retrains. The curves are conceptual; no measured data are plotted.
ILLUSTRATION: Think of a lab analyzer. You run QC daily so calibration drift is caught before patient results go wrong; you do not wait for a clinician to notice odd values. Output review is waiting for the clinician; drift monitoring is daily QC.
FDA PARALLEL: Predetermined change control plans (final guidance, Dec 2024) let device makers pre-specify which model updates are allowed and what testing each must pass. The diamond on the teal line is that idea applied inside your organization.
REAL-WORLD PATTERN: Ambient scribes and message drafters usually sit on a vendor’s foundation model. If the vendor changes that model, your validated tool is now a different tool, whether or not the product name changed.
WHAT IMPLEMENTATION BUYS YOU: earlier detection (leading indicators instead of lagging ones); a clean audit trail (which version produced which output); and a defensible answer when a surveyor or plaintiff asks when you knew.
DECISION ASK: For every deployed model, require (1) advance change notification, (2) a revalidation gate before any new version reaches patients, and (3) at least one leading indicator, such as input or embedding drift, in addition to output review.
Requirements (1 of 2)
Requirements that differ by tier
Requirement SLM MLM LLM CHAI model card Authored in full by you Full if open weights Vendor’s; site logs gaps Validation MedHELM subsets, local set Same, plus subgroups Local data or vendor only Abstention Trained via preference pairs Trained or prompted Prompted; clinician review Change control Pinned, hashed; card per version Pinned; re-test after quantizing Revalidate on vendor update Monitoring Outputs, logits, embeddings Internals if self-hosted Output level only Privacy / Part 2 On-prem or on-device On-prem or private cloud BAA; Part 2 consent limits
MLM column interpolated · Healthcare LM Governance · Oct 2026
16
KEY MESSAGE: The smaller and more self-hosted the model, the more evidence you produce yourself. The larger and more vendor-hosted, the more you document gaps and revalidate on changes you do not control.
WALK THE ROWS: Model card: author it versus request it. Validation: MedHELM subsets plus a local set for SLMs; add subgroup analysis at MLM scale; for LLMs at small sites, vendor evidence may be all you have. Abstention: trained into SLMs through preference pairs; prompted elsewhere and tested either way. Change control: pinned and hashed versus revalidate on every vendor update. Monitoring: self-hosted models allow internal signals (logits, embeddings); vendor APIs allow output review only. Privacy: on-device or on-prem versus a BAA, with Part 2 consent limits for SUD data.
ILLUSTRATION: Owning the recipe versus reading the label.
WORKED EXAMPLE: A clinic adopting a vendor LLM message assistant gets the card from the vendor, documents the gaps, reviews a 50-message chart sample monthly, and checks BAA and Part 2 consent coverage. The same clinic running a tuned 7B SLM authors its own card, versions it per release, runs a drift dashboard, and never sends PHI off-site. The second path is more work up front and far less negotiation later.
CAVEAT: The MLM column is interpolated (see the model-tiers slide).
DECISION ASK: Adopt this table as the intake checklist, applied by tier.
Requirements (2 of 2)
Requirements that apply to every tier
Intended use and scope
Name the target MedHELM subcategories and out-of-scope uses before any evaluation.
NIST 600-1 worksheet
Confabulation, human-AI configuration, data privacy and value chain, each with controls.
Device-boundary memo
Clinician- vs. patient-facing, tested against the four FDA 2026 CDS criteria.
Safety-event reporting
Route events into the site’s voluntary, blinded RUAIH reporting process.
Healthcare LM Governance · SLM Sprint · Oct 2026
17
KEY MESSAGE: Four requirements apply regardless of model size. Think of them as the universal precautions of clinical AI.
INTENDED USE AND SCOPE: Write it before evaluation. It sets which MedHELM subcategories you test and what goes on the out-of-scope list. Most governance failures start as unscoped reuse.
NIST 600-1 WORKSHEET: Four rows (confabulation, human-AI configuration, data privacy, value chain), each with a named control and an owner.
DEVICE-BOUNDARY MEMO: One page answering three questions. Is the user a clinician or a patient? Can the clinician independently review the basis for the output? Is the decision time-critical? This is where the January 2026 FDA changes matter: single recommendations may now be acceptable for clinician-facing tools, while patient-facing tools get no new relief.
SAFETY-EVENT REPORTING: Route AI events into the existing voluntary, blinded reporting process, so AI incidents reach the same quality team that sees medication errors. Do not build a separate AI incident silo.
ILLUSTRATION: Universal precautions apply to every patient regardless of what you believe the risk is. These four apply to every model regardless of what the vendor says about it.
DECISION ASK: Make these four a gate in the project-approval workflow, so nothing reaches a pilot without them.
Implementation · What post-training changes
Preference tuning moves the model toward the safe answer
Healthcare LM Governance · SLM Sprint · Oct 2026
18
KEY MESSAGE: Post-training is how a small model earns its scope. Preference tuning shifts probability toward the safe, useful answer and away from both confident errors and unhelpful refusals.
HOW TO PRESENT THE ANIMATION: One patient message, three candidate replies. The stage chips step from the base model, to after supervised fine-tuning (SFT), to after direct preference optimization (DPO), and the bars move with them. Stress that the bar lengths are illustrative, not measured.
WHAT EACH STAGE DOES: SFT teaches the format and tone of good replies from curated examples. DPO then trains on pairs, a chosen and a rejected reply to the same prompt, increasing the relative likelihood of the chosen one. A frozen reference model and the β parameter keep the tuned model from drifting too far from what it already knew. DPO needs no separate reward model, which is why it fits a one-GPU budget with LoRA adapters (the sprint’s Mistral-7B setup uses r=16, α=32).
WHY TWO REJECTED REPLIES: Safety tuning that only punishes dangerous advice tends to over-teach refusal. Pairing the good reply against a bare refusal as well keeps the model helpful. Measure helpfulness alongside safety, or the model will learn that silence is safest.
ILLUSTRATION: Preference pairs work like precepting a resident: you don’t just say what not to do, you model the better answer next to the worse one, case after case.
GOVERNANCE TIE: Each preference pair is a documented decision about what behavior the organization rewards. A committee can read the rubric and a sample of pairs. That is evidence a vendor LLM cannot hand you, and it maps directly onto the CHAI model card’s intended-use and known-risk fields.
PERFORMANCE GAINED BY IMPLEMENTING IT: a narrower, more predictable output distribution on the target task; explicit escalation behavior that reviewers can audit; and lower serving cost than a frontier model for the same high-volume task.
CAVEAT: A model tuned on one task’s pairs is only validated for that task. Re-tag pairs by MedHELM subcategory so scope stays explicit.
DECISION ASK: Treat the clinician-written preference rubric as a governed artifact, approved by the committee and versioned with the model.
Sprint realignment
Each phase ends in a governance artifact
Phase Answers to Revised deliverable 1 · DPO MedHELM Patient Communication Pairs tagged by subcategory; MedHELM rubric 2 · PEFT NIST environmental; value chain Adds license, adapter provenance, GPU-hours 3 · Evaluation MedHELM primary; MedQA floor MedHELM subsets; jury vs. you on 50 items 4 · Production FDA 2026 CDS; NIST confabulation vLLM benchmark; device-boundary memo 5–6 · Portfolio CHAI card; RUAIH monitoring CHAI JSON card on HF Hub; drift monitor 7 · Interview RUAIH five areas; HTI-5 Talk track on supporting certification
Target: ambulatory and behavioral health, MedHELM Patient Communication.
Healthcare LM Governance · SLM Sprint · Oct 2026
19
KEY MESSAGE: Every sprint phase now ends in something a governance committee would ask for, so the portfolio doubles as a governance evidence package.
WALK THE PHASES: Phase 1, DPO: preference pairs tagged by MedHELM subcategory, with a chosen/rejected rubric written to MedHELM criteria. Phase 2, PEFT: the comparison table adds base-model license, adapter provenance and GPU-hours (NIST value-chain and environmental rows). Phase 3, evaluation: MedHELM subsets plus clinician–jury agreement on 50 items. Phase 4, production: a vLLM latency benchmark plus the device-boundary memo. Phases 5–6, portfolio: a CHAI card in JSON shipped with the model on the Hugging Face Hub, and a drift and subgroup monitor. Phase 7: a talk track on supporting a site’s RUAIH certification without being certified.
STATUS (last recorded 8 Sep 2026): Environment complete (Python 3.13, torch 2.14, transformers 5.4, trl, peft). Mistral-7B with LoRA (r=16, α=32) DPO script built on 100 UltraFeedback pairs; DPO paper reading in progress. Phases 2 through 7 not started.
ILLUSTRATION: Residency logic. Each rotation ends in a signed evaluation, not just logged hours.
WHY IT MATTERS FOR DECISION-MAKERS: A candidate or vendor who can hand over these artifacts has already done the work your committee would otherwise have to demand.
DECISION ASK: Measure progress by artifacts produced, not weeks elapsed.
Draft a CHAI model card for the DPO checkpoint
Confirm trl works with transformers 5.x in one training step
Write the PMSG device-boundary memo
Re-tag DPO pairs by MedHELM subcategory
Add a NIST 600-1 worksheet for the four risks
Run one MedHELM patient-communication benchmark, base vs. DPO
Read RUAIH guidance and one CHAI playbook; write a half-page
Score 50 outputs; report clinician–jury agreement
Check whether HTI-5 has been finalized
Choose portfolio A (SUD counseling DPO) or B (chronic pain + wearable)
Healthcare LM Governance · SLM Sprint · Oct 2026
20
KEY MESSAGE: Eleven concrete steps. The first three unblock everything else.
PRIORITY ORDER: (1) Log progress since 8 Sep so the plan restarts from the right week. (2) Confirm trl works with transformers 5.x by running one training step end to end; trl 0.8 predates transformers 5.x, and the current DPOTrainer takes a DPOConfig. (3) Re-tag the preference pairs by MedHELM subcategory, so scope is explicit from the start.
THEN, IN PARALLEL: Install crfm-helm and benchmark base versus DPO on one patient-communication task. Score 50 outputs yourself and report clinician–jury agreement. Draft the CHAI Applied Model Card from the mc-schema repository on GitHub. Write the PMSG device-boundary memo. Fill in the four-row NIST worksheet. Read the RUAIH guidance and one CHAI playbook and write a half-page on supporting certification.
KEEP CHECKING: HTI-5 has been at OMB since 28 Jul 2026 and was not final as of mid-October 2026. Recheck before each presentation and update the timeline slide.
DECIDE: Portfolio option A (SUD counseling DPO) or B (chronic pain plus wearable). Option A sits squarely in the Part 2 and state therapy-chatbot rules covered on the care-settings slides; option B touches device-boundary questions if wearable data drive recommendations.
DECISION ASK: Assign an owner and a date to each item before the next meeting.
MedHELM (arXiv 2505.23802; Nature Medicine)
Polsinelli: TJC/CHAI RUAIH guidance
Fierce Healthcare: RUAIH guidance and certification
Hackensack Meridian Health: first RUAIH certification
CHAI Applied Model Card registry and mc-schema
Anaconda: NIST AI 600-1 guide
arXiv 2604.22768 and 2601.03266 (on-prem and on-device LLMs)
Covington and Drummond Group: HTI-5 proposed rule
WEDI and Live Compliance: HTI-5 at OMB review (Jul–Oct 2026)
Pacific AI: HTI-5 and Section 1557 92.210
Covington: FDA 2026 CDS guidance takeaways
Ropes & Gray: EO 14365 and the March 2026 framework
McDermott: Colorado SB 26-189
Becker’s: five states restrict AI therapy chatbots (2026)
Gibson Dunn and Lewis Silkin: EU Digital Omnibus on AI
Obermeyer et al., Science, 2019
Wong et al., JAMA Internal Medicine, 2021
Texas Attorney General: Pieces Technologies settlement, 2024
FDA: final PCCP guidance, Dec 2024
Darrell A. Hall, MD, MSCIS · dah512@edmhs.us · Status current to 20 Oct 2026
21
Column one holds the sources from the 14 October 2026 taxonomy alignment document. Columns two and three add the status checks and historical examples used in the speaker notes, current to 20 October 2026. Two items are moving and should be rechecked before each presentation: HTI-5 (under OMB review since 28 July 2026 and not yet final in mid-October) and federal preemption of state AI laws (EO 14365 and the March 2026 legislative recommendations are not binding law; state laws remain in effect until Congress or the courts act).