Bettering Bearings LLC

Bettering Bearings, LLC submits the attached comment on FDA’s discussion paper, *Considerations for the Regulation of Generative AI-Enabled Medical Devices*.

Our central recommendation is that FDA distinguish demonstrated AI task performance from the conditions necessary for responsible clinical judgment. Accurate outputs, successful benchmark performance, and nominal human oversight do not, by themselves, establish that a device can be used safely and responsibly in its intended clinical setting.

We recommend that FDA consider five complementary evaluation requirements:

1. **Risk and action-directiveness:** Assess how outputs influence actual decisions, including when ostensibly informational responses become functionally directive through workflow placement, personalization, urgency, or repeated reinforcement.

2. **Clinical competency and judgment preservation:** Evaluate whether clinicians and patients can recognize uncertainty, identify the limits of their competence, seek appropriate consultation, and contest or reject recommendations under realistic conditions.

3. **Escalation and intervention:** Test under-escalation and over-escalation separately. For consequential functions, establish whether authorized humans possess the information, time, authority, and practical capacity to interrupt or correct device actions.

4. **Agentic functions:** Distinguish authorization to plan from authorization to execute. Evaluate delegation boundaries, tool use, handoffs, stopping behavior, and recovery from errors across multi-step clinical workflows.

5. **Postmarket accountability:** Monitor changes in reliance, clinical workflows, model behavior, and intervention capacity. Define reassessment triggers when system modifications materially affect the original evidence supporting safe use.

We propose a Judgment-Preservation Assurance Record connecting intended use, supporting evidence, known limitations, delegated authority, intervention tests, escalation conditions, and postmarket reassessment.

These recommendations are offered as candidate evaluation criteria requiring further empirical validation, not as claims that particular thresholds have already been established.

The attached memorandum provides detailed recommendations organized around FDA’s discussion questions, including Questions 1–6 and 26.

Bettering Bearings, LLC

Clinical Competency Is Not Yet Clinical Judgment
Bettering Bearings, LLC | FDA docket FDA-2026-N-7874 | Public comment | 19 November 2026
Purpose and scope
Bettering Bearings, LLC offers focused comments on FDA's discussion paper, not a proposed
binding rule. The contribution addresses device-software functions within FDA's remit and does
not suggest regulating AI models as such. The central recommendation is to distinguish
performance evidence from the conditions under which clinicians and patients can exercise
informed, timely, accountable judgment.
Questions 1–2 — risk dimensions and action-directing outputs
Retain FDA's proposed axes of activity and consequence as an organizing heuristic, but use a
structured modifier record for (a) time available for independent review, (b) reversibility of
downstream action, (c) source traceability, (d) user capacity to detect error, and (e) the extent to
which a recommendation induces action despite nominally informational wording. These
modifiers should be tied to intended use and deployment context, not treated as a third
undifferentiated risk score.
Test action-directiveness using realistic user behavior, not disclaimer language alone. A model's
apparent suggestion may be functionally directive where personalized urgency, repeated
reinforcement, workflow placement, or authority cues make reliance likely. Record when and
how informational outputs cross into action-directing behavior.
Questions 3–5 — patients, expertise, and conversational drift
For patient-facing functions, avoid assuming either incapacity or competence from professional
status alone. Evaluate whether a user can recognize the output's limitations, locate an appropriate
clinical escalation route, and act within the available time. The adequacy of these safeguards
should be tested with representative users and accessibility needs.
For generalist clinicians using specialist-domain tools, test not only answer accuracy but
recognition of the boundaries of the clinician's own competence, ability to verify claims, and
escalation to specialty care. A useful benchmark includes cases where correct action is to defer,
seek consultation, or decline a recommendation.
For multi-turn devices, use trajectory-based tests that identify transitions from explanation to
recommendation to action initiation. Record the earliest point at which risk changes, whether the
system recognizes that transition, and whether its safeguards adapt in time.
Question 6 — escalation tradeoffs
Assess under-escalation and over-escalation separately before considering aggregate
performance. A single composite accuracy figure can conceal unacceptable failure on rare but
consequential presentations. Define context-specific thresholds and report uncertainty, patient
burden, unnecessary interventions, and delay-sensitive harm transparently.
Competency-based premarket evaluation
Clinical competency is an instructive analogy, but device competence is not clinical professional
judgment or moral agency. A proposed assurance case should link each intended-use claim to (1)
test evidence, (2) known boundary conditions, (3) authorized human responsibility, (4)
intervention and override mechanisms, and (5) triggers for re-evaluation. Nonclinical benchmarks
and clinical confirmation should test failures of warranted deference, not just successful task
completion.
Include counterfactual and adversarial cases in which the safest system behavior is to expose
uncertainty, ask for missing context, refrain from directing care, or escalate to an appropriately
qualified person. Evaluate whether clinician users can actually detect and correct device error
under realistic workload and time pressure.
Postmarket monitoring and changes
Monitoring should detect drift in reliance and workflow effects as well as changes in model
outputs. Proposed indicators include unreviewed adoption of recommendations, override rates
with reason codes, delayed escalation, clinician deskilling signals, failures of source retrieval, and
recurrence of previously identified defeaters. These indicators are proposed research and
assurance measures, not claims that validated thresholds already exist.
Require an explicit revision trigger when a foundation model, prompt, retrieval source,
orchestration layer, or clinical interface changes in a way that could alter action-directiveness or
reviewability. Changes should be assessed against the intended use and the evidence supporting
the original assurance case.
Question 26 — agentic medical-device functions
For multi-step agentic functions, map each authorized action, tool, data source, and clinical
boundary to a responsible human or institution. Distinguish permission to plan from permission
to execute. Test stop, pause, rollback, and handoff behavior at each clinically consequential step,
including failed tool calls and conflicting instructions.
A device should not be regarded as adequately supervised solely because a clinician is nominally
'in the loop.' The relevant test is whether the clinician has sufficient information, time, authority,
and practical ability to interrupt or correct the action before material harm. Where such capacity
is absent, the deployment should be assessed as more autonomous for risk purposes.
Suggested evaluation artifact
A concise 'Judgment-Preservation Assurance Record' could contain: intended-use and activity
classification; evidence/provenance ledger; known failure modes; delegation and authority map;
review-time budget; escalation and defeater triggers; intervention test results; and postmarket reevaluation conditions. FDA could invite stakeholder evidence on which elements are measurable
and proportionate across device classes.
Sources
FDA, 'Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion
Paper and Request for Feedback' (August 2026), https://www.fda.gov/media/194242/download;
FDA consultation page: https://www.fda.gov/medical-devices/digital-health-centerexcellence/considerations-regulation-generative-ai-enabled-medical-devices-discussion-paperand-request.