Limbic (May Lee)

See attached file for Limbic’s comments

Limbic Comments on Docket No. FDA-2026-N-7874

Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and
Request for Feedback

Submitted by May Lee, Chief Compliance and Regulatory Officer, Limbic, on October 19, 2026

Limbic appreciates the opportunity to comment on the Center for Devices and Radiological Health's
discussion paper, Considerations for the Regulation of Generative AI-Enabled Medical Devices (August
2026). We commend CDRH for engaging stakeholders early on these questions, and for proposing an
approach grounded in risk, the total product life cycle, and least burdensome principles.
Limbic develops clinical AI for behavioral health. Our product is Unpacked, a generative AI-enabled
conversational therapy device participating in FDA's TEMPO pilot. Our comments draw on our experience
developing, deploying, monitoring, and changing patient-facing conversational AI under regulatory oversight
in the United States and the United Kingdom.

Our key recommendations are as follows:
1.​ Evaluate the device as a system. We support CDRH's proposal to evaluate the final user-facing
device as configured for deployment. We ask CDRH to apply that principle consistently across risk
assessment, acceptance criteria, change control, and postmarket monitoring. (Overarching
comments; Q1, Q6, Q25)
2.​ Treat benchmarking and monitoring as one continuous process. The premarket benchmark
should serve as the baseline for re-benchmarking, staged rollout, and postmarket monitoring
throughout the device's life. (Overarching comments; Q7, Q18, Q19)
3.​ Refine the two-axis risk framework. We recommend adding the device's action envelope and
layered safeguards as modifiers. We also ask CDRH to address treatment-delivery devices, where
directive content is inherent to evidence-based care. (Q1, Q2, Q8, Q26)
4.​ Preserve sponsor flexibility in benchmark design, with safeguards for validity. Device-specific
benchmarks should be supported by clinician-labeled scoring, held-out test sets, and postmarket
evidence that benchmark performance predicts real-world performance. (Q9, Q10, Q13)
5.​ Adapt PCCPs for GenAI-enabled devices. We recommend allowing a PCCP to prespecify the
verification gate that each modification must pass, including changes to third-party foundation
models. (Q22, Q23, Q24)
Overarching comments

Limbic supports CDRH's risk-based, total product life cycle approach. We also support its proposal to
evaluate the final user-facing device as configured for deployment (Section V.A, p. 11). We ask CDRH to
apply that principle consistently across risk assessment, premarket evaluation, change control, and
postmarket monitoring. Two points run through our responses.

First, the device system is the appropriate unit of evaluation. In conversational GenAI-enabled devices,
safety rarely depends on a single component. It comes from layered controls: real-time detection, gating of
every output before it reaches the user, restrictions on the actions the device can take, automated review of
completed sessions, and clinician follow-up. Risk, acceptance criteria, and change control should be
assessed against the performance of that whole system. Thresholds applied to individual components
would penalize layered designs and would restrict improvements to modifications of existing components.

Second, manufacturer-defined benchmarking and postmarket monitoring should operate as a single
continuous, iterative process that extends from premarket evaluation through the life of the device.
GenAI-enabled devices can be changed in controlled increments, and the effect of each change can be
measured in real-world use. We recommend CDRH recognize a cycle in which each change passes a
pre-release gate proportionate to its potential impact, is deployed to a limited population under monitoring,
and is expanded or rolled back against prespecified criteria, with monitoring findings feeding back into the
benchmark. This approach complements premarket assessment by placing greater weight on real-world
performance.

Section IV: Assessment of risk

Question 1: Dimensions of the two-axis risk framework

FDA question Does the two-axis risk framework, organized around device activity
and the consequence of relying on an incorrect output, appropriately
capture the dimensions most relevant to the risk of a GenAI-enabled
software function? If there are additional dimensions, such as the
reversibility of a resulting action, the availability of downstream
safeguards, the time pressure of the deployment setting, or the
traceability of the output (i.e., to primary source materials), that should
be represented in a risk framework, please describe and provide
examples of how they should be represented.

Position Support with modifications

Recommendation to Limbic supports the two-axis framework as a heuristic and
FDA recommends two additional modifiers.
The first is the device's action envelope: the set of actions the device
is technically enabled to take, and the controls gating them. A device
that can only schedule a session or connect a patient to a crisis line
differs materially from one that can prescribe, even if both converse
freely.
The second is layered safeguards and reversibility. In conversational
devices, no single safety component is perfect. Safety comes from
layers such as real-time detection, automated post-session review,
and clinician review. Risk should therefore be assessed for the system
as a whole, taking these safeguards into account.

Citations Discussion paper, Section IV and Figure 1, pp. 5–6.
Question 2: Directiveness continuum

FDA question CDRH seeks input on how to account for the spectrum of
GenAI-enabled informational functions that vary in the degree to which
they direct a user to a particular action (i.e., between "non-directive"
and "action-directing"). What characteristics of an output, such as its
wording, specificity, personalization, or context, could be considered
as modifiers of the risk associated with an informational function, after
accounting for the device's overall functionality and intended use?
What additional information would provide manufacturers with
sufficient clarity and predictability around risk assessment for
informational functions while recognizing that directiveness may exist
along a continuum rather than as a binary distinction?

Position Support with modifications

Recommendation to The examples in Section IV involve discrete clinical decisions such as
FDA medication dosing and triage. GenAI is also being used to deliver
treatment, particularly in mental health, where the conversation itself is
the intervention (see, e.g., Rollwage et al., 2026). Limbic authors
contributed to that study. In manualized psychotherapy such as
cognitive behavioral therapy, directive content is inherent to the
treatment; examples include behavioral activation tasks and cognitive
restructuring prompts. Applied without adjustment, the directiveness
continuum would place nearly all such output at the action-directing
end.
We recommend that CDRH address treatment-delivery devices
explicitly, including the additional guardrails they require. We also
recommend that CDRH recognize objective modifiers of directiveness
risk: adherence to an established, evidence-based treatment protocol;
the device's action envelope; and the presence of clinician oversight.
Objective modifiers would give manufacturers the predictability this
question seeks.

Citations Discussion paper, Section IV, p. 7. Rollwage M, McFadyen J, Juchems
K, et al. A cognitive layer architecture to support large-language model
performance in psychotherapy interactions. Nat Med.
2026;32:1717–1725. doi:10.1038/s41591-026-04278-w.

Question 3: Patient-facing vs. HCP-facing functions

FDA question CDRH seeks input on whether and when a GenAI-enabled function
that results in delivery of clinical information to patients, as opposed to
HCPs, could present different or higher risks, while also recognizing
the potential benefits associated with improved patient empowerment,
engagement, and access to clinical information. What device
characteristics, output features, or safeguards might mitigate risks that
could arise when a user lacks the domain knowledge to independently
evaluate an output, without unnecessarily underestimating patient
capability?

Position Support with modifications

Recommendation to Limbic does not support an automatic upward shift on the
FDA consequences axis for patient-facing functions. Mitigations should
count against any such shift. These include validated escalation
 pathways, clinician oversight, and testing under benchmarking
element E.4 that shows intended users understand outputs and know
when to seek help.

Citations Discussion paper, Section IV, p. 8; Appendix A, E.4, p. 25.

Question 4: Generalist vs. specialist HCPs

FDA question CDRH seeks input on whether and how the distinction between
generalist and specialist physicians might inform the assessment of
risk for HCP-facing GenAI-enabled functions. Under what
circumstances, if any, might risk be affected when an HCP who lacks
the relevant clinical specialist knowledge receives information that falls
within a specialist area of practice? What device characteristics or
safeguards might mitigate such risks?

Position Support with modifications

Recommendation to Limbic agrees with CDRH that the expertise of the intended user can
FDA affect risk. Where safe use of an output depends on specialist
contextualization, the role of specialist expertise should be reflected in
risk assessment and evaluation. We recommend that CDRH frame
this as a question of whether the intended user can independently
evaluate the output, consistent with footnote 13 of the discussion
paper.
Risk is most likely to increase when three conditions coincide: (1) the
output falls within a specialist domain; (2) acting on an incorrect output
carries moderate or severe consequences; and (3) the intended user
lacks the training to recognize that the output is wrong. In behavioral
health, for example, a function that recommends starting an
antidepressant without adequately screening for bipolar disorder could
lead a generalist to an action whose risks a psychiatrist would be more
likely to recognize. That risk is best expressed as a shift on the
consequences axis of the existing framework.
Risk should not be presumed higher simply because a function
delivers specialist-level knowledge to generalists. In domains where
generalists already deliver much of the care, such functions can
extend access to specialist expertise, a benefit CDRH recognizes (p.
8). Penalizing them by default would discourage tools that help close
access gaps.
We recommend that CDRH recognize the following mitigations.
Labeling. Labeling that defines the intended user and any required
competencies.
Scope recognition. Outputs that identify when a question exceeds
the intended user's scope and recommend specialist consultation or
referral, evaluated under benchmarking elements S.2 and S.3.
Traceability. Transparency about the basis for an output, so the user
can check it against clinical guidelines.
Specialist oversight within the system. Specialist review of flagged
cases or defined escalation pathways to specialists.
Evaluation matched to the intended user. Where safe use depends
on specialist expertise, benchmarking adjudicators should have
qualifications in that specialty, consistent with Section V.B.2. Where a
function is intended to support generalists, clinical confirmation should
compare the generalist working with the device against the generalist
 working without it, consistent with the human-AI team and
absence-of-device comparators discussed in Section V.D.1 and Q15.

Citations Discussion paper, Section IV, p. 8 and n. 13; Section V.B.2, p. 14;
Section V.D.1, p. 16. FD&C Act section 520(o)(1)(E). FDA, Clinical
Decision Support Software (January 2026).

Question 5: Multi-turn conversational devices

FDA question For multi-turn conversational GenAI-enabled devices that may migrate
from providing "non-directive" information to "action-directing"
information over the course of an exchange, how should risk be
assessed across realistic conversational trajectories? How could the
intended use of such a device be characterized when its behavior is
emergent across a conversation?

Position Support with modifications

Recommendation to Limbic supports trajectory-level testing. Intended use should continue
FDA to be determined by the manufacturer's objective intent (21 CFR
801.4). User attempts to steer a device beyond that intended use are
properly addressed through testing under benchmarking element S.2.
Intended use can also be bounded architecturally. Where every output
passes through a validated gating layer before reaching the user,
whichever component generated it, and the device's available actions
are restricted, trajectory testing should focus on whether those
controls hold across long and adversarial conversations. Monitoring of
real long conversations after deployment then confirms that the
controls continue to hold.

Citations Discussion paper, Section IV, p. 8; Appendix A, S.2, p. 24; 21 CFR
801.4.

Question 6: Care escalation errors

FDA question For GenAI-enabled care escalation functions, CDRH is considering
how the evaluation may account for both under-escalation and
over-escalation. How could manufacturers characterize and weigh
these two directions of error, given that they may not be
commensurable and that acceptable trade-offs may vary by clinical
context?

Position Support with modifications

Recommendation to Limbic supports evaluating both directions of error and notes that the
FDA two carry different weight. In behavioral health, under-escalation can
be fatal, for example when an indication of suicide risk is missed, so it
generally warrants greater weight. Over-escalation also causes real
harm: unnecessary welfare checks or police involvement, patient
distress, and disengagement from care that can reduce appropriate
help-seeking later.
The appropriate weighting depends on clinical context. We therefore
recommend that sponsors prespecify and justify their weighting and
their sensitivity and specificity criteria, and that CDRH evaluate those
criteria for each device. Where criteria apply, they should apply to the
 escalation performance of the whole system (in-session detection,
post-session review, and clinician follow-up).
We also ask CDRH to acknowledge how device escalation behavior
interacts with state mandatory-reporting and duty-to-warn obligations.

Citations Discussion paper, Section IV, pp. 8–9; Appendix A, S.1, p. 24.

Section V: Competency-based premarket evaluation

Question 7: Competency-based approach overall

FDA question Is the competency-based approach described above, i.e., device
benchmarking followed by clinical confirmation, a useful and
appropriate framework for evaluating GenAI-enabled devices?

Position Support with modifications

Recommendation to
FDA Limbic supports the competency-based approach of device
benchmarking followed by clinical confirmation. It is a practical
framework for devices whose inputs and outputs are too varied for
exhaustive testing. We particularly support CDRH’s decision to
evaluate the final user-facing device as configured for deployment
(Section V.A). This matches how safety is achieved in conversational
devices, where it comes from layered controls across the whole
system.

We recommend three refinements.

Treat the competency assessment as the start of a continuous
process. The premarket benchmark should serve as the baseline for
postmarket re-benchmarking, change control, and monitoring.
Benchmarking, clinical confirmation, and postmarket evidence then
form one connected body of evidence across the device’s life cycle.

Preserve sponsor flexibility in benchmark design, with
safeguards for validity. Device-specific benchmarks are necessary
to test scenarios specific to an intended use. Construct validity can be
shown through clinician-labeled scoring, held-out and refreshed test
sets, and postmarket evidence that benchmark performance tracks
real-world performance.

Recognize where the clinician analogy needs adaptation. Human
clinicians face professional, legal, and reputational consequences
when they fall short. For devices, the equivalent accountability comes
from enforceable change control, continuous monitoring with
prespecified rollback criteria, and manufacturer responsibility for the
whole system. The framework should state these mechanisms
explicitly.

We also ask that the approach be applied consistent with the least
burdensome principles CDRH cites (p. 1). Evidence requirements
should scale with risk, as discussed in Q8.

Citations Discussion paper, Section V.A, pp. 10–11.

Question 8: Using the risk framework to set evidence levels
FDA question How might the two-axis risk framework described in Section IV be
considered within the competency-based approach described above
to help determine the level of evidence needed for a premarket
submission?

Position Support with modifications

Recommendation to
FDA Limbic supports using the two-axis framework to scale the nature and
amount of premarket evidence. We recommend CDRH apply it to four
aspects of the evidence package:

●​ the benchmarking elements selected
●​ the range and difficulty of test conditions
●​ the type and rigor of clinical confirmation
●​ the strength of the postmarket monitoring program

The framework’s own modifiers should carry through to evidence
requirements:

Action envelope and layered safeguards. A device whose external
actions are narrow and gated by deterministic controls warrants less
premarket evidence than one that can take irreversible clinical actions,
even if both converse freely.

Mitigations for patient-facing functions. Where a patient-facing
function has validated escalation pathways, clinician oversight, and
demonstrated user comprehension (E.4), those mitigations should
offset any upward shift in evidence requirements.

The intended user’s ability to evaluate outputs. Evidence
requirements should reflect whether the intended user can
independently evaluate the output. Adjudicators and comparators
should be matched to that user.

Sponsor-justified error weighting. For functions with escalation
decisions, sponsors should prespecify how they weight under- and
over-escalation and set acceptance criteria at the system level.

Postmarket monitoring. The framework should also govern the
balance between premarket and postmarket evidence. Where a
sponsor can deploy changes in controlled increments under
monitoring, with rollback, a device may justify less premarket evidence
than its position on the axes alone would suggest.

To give sponsors predictability, we recommend CDRH publish
illustrative examples showing how devices at different points on the
framework would map to benchmarking scope and clinical
confirmation approaches. Each sponsor would still justify its own
evidence plan.

Citations Discussion paper, Section V.A, p. 11.

Question 9: Benchmarking elements

FDA question Would a benchmarking structure such as the one described above be
likely to provide adequate evidence of clinical knowledge, analytic
 capabilities, safety behavior, communication, and generalizability to
support a reasonable assurance of safety and effectiveness? Are
there elements that are missing, redundant, or inappropriately
categorized? Are there externally developed standards that could be
leveraged?

Position Support with modifications

Recommendation to Limbic supports the benchmarking structure and the flexibility for
FDA sponsors to propose device-specific benchmarks, which is necessary
to test scenarios specific to an intended use. We ask CDRH to provide
clarity on evaluation methods while preserving that flexibility, and we
recommend the following refinements.
Reliability (R.1). As drafted, R.1 treats any variation in safety-critical
behavior as a failure, a standard no probabilistic system can meet. We
recommend a statistical definition instead: consistency rates with
confidence bounds, evaluated against prespecified acceptance
criteria.
Information gathering (E.2). E.2 applies beyond diagnosis. In
treatment-delivery devices, assessment, case formulation, and risk
exploration are all information-gathering tasks, and the element should
be framed to include them.
Continuity across sessions (proposed addition). Treatment runs
over many sessions, but the current elements address single
conversations. Correct use of prior-session information, such as
history, safety plans, and progress, should be benchmarked.
Over-reliance and dependency (proposed addition under E.4 or
Safety). In mental health, over-reliance on a GenAI-enabled device is
a safety risk. For example, reassurance-seeking is a core feature of
obsessive-compulsive disorder, and excessive device use can
reinforce it. Usage quantity alone is a poor measure, because
sustained engagement is also a treatment goal. Measures should
combine content signals, such as repeated reassurance-seeking, with
usage patterns, and sponsors should define and justify them.
Speech input (R.2). For voice devices, speech recognition is part of
the system. Evaluation should test whether safety-critical behavior,
such as risk detection and escalation, holds across accents, dialects,
and noisy audio. A standalone transcription accuracy requirement
would be unworkable, because that metric cannot be measured in
end-to-end speech models, where speech processing and response
generation are integrated in a single model. Such a requirement would
block adoption of those architectures.

Citations Discussion paper, Section V.B.1, pp. 12–13; Appendix A, pp. 24–26.

Question 10: Benchmark validity

FDA question CDRH recognizes that publicly available benchmarking assets may be
subject to data contamination, saturation, and limited real-world
representativeness. How should a sponsor establish that performance
on a given benchmark predicts safe and effective real-world behavior
for the device's intended use? What evidence should support the
construct validity of a benchmark used to gate device evaluation, and
what role should sponsor-developed benchmarks play given potential
concerns around independence and optimization to the test?
Position Support with modifications

Recommendation to Limbic supports sponsor-developed benchmarks, which are necessary
FDA to test scenarios specific to the intended use, provided construct
validity is shown by three things: (1) scoring grounded in labels from
qualified clinicians; (2) held-out test sets that are never used in
development or optimization and are refreshed over time; and (3)
postmarket evidence that benchmark performance tracks real-world
performance measured through monitoring.
Where a device is optimized against automated judges (for example,
through automated prompt optimization or reinforcement learning), the
judges used for evaluation and monitoring should be separate from
those used for optimization, and validated against human labels.

Citations Discussion paper, Section V.B, p. 12.

Question 13: Suitability of synthetic data

FDA question For which clinical domains, device functions, or subpopulations is
synthetic data particularly well-suited, or particularly inadequate, as a
supplement to real-world evidence? What safeguards would mitigate
the risk that synthetic data generated by models of the same class as
the device under evaluation reproduces the very performance gaps
the evaluation is intended to detect, particularly for underrepresented
subgroups?

Position Support with modifications

Recommendation to Synthetic and simulated data are well suited to regression testing and
FDA to covering known safety-critical scenarios at scale, including
multi-turn conversations. They are valuable as part of a pre-release
gate.
They are poorly suited to discovering new failure modes. In mental
health, those often emerge over long interactions shaped by real
patient behavior that simulated patients do not reproduce. Synthetic
evaluation should therefore be complemented by monitoring of
real-world interactions and by targeted human testing for
underrepresented subgroups. Both the simulations and the judges
scoring them should be anchored to clinician-labeled real data.
Performance estimates from synthetic and real-world data should be
reported separately unless calibration between them has been
demonstrated. Statistically meaningful real-world estimates come most
naturally from monitoring all production interactions.

Citations Discussion paper, Section V.C, p. 16.

Question 14: Performance comparators

FDA question For open-ended device outputs, how should performance comparators
and acceptance criteria be selected? When a panel of qualified
clinicians serves as the comparator, how should the applicable
standard (for example, the standard of care versus the performance of
a median clinician in practice) be defined and justified? Should
generalist or specialist physicians be used as a performance
 standard? When should human-AI team performance, rather than the
device operating alone, serve as the basis for evaluation?

Position Support with modifications

Recommendation to Comparators and acceptance criteria should match the device's
FDA intended use and intended user (see Q4). Where safe use of an
output depends on specialist expertise, specialists should serve as the
comparators and adjudicators. Where a device is intended to support
generalists or to work alongside a clinician, the primary comparison
should be the clinician working with the device against the clinician
working without it, together with the absence-of-device comparator
discussed in Q15. Comparison to the device operating alone is most
appropriate for functions intended to operate without clinician
involvement.
Where validated fidelity or competence instruments exist for the
clinical activity (for example, therapist competence rating scales in
cognitive behavioral therapy), scoring rubrics should be anchored to
those instruments. An undefined "median clinician" standard would be
difficult to apply consistently. We also note that "standard of care" has
a specific meaning in tort law. We recommend that CDRH define the
performance standard explicitly so that device evaluation does not
import liability concepts.

Citations Discussion paper, Section V.D.1, p. 16; Section IV, p. 8.

Question 15: Absence-of-device comparator

FDA question Are there ways in which performance might be assessed relative to
the care, technology, or course of action likely to occur in the absence
of the device, rather than to the comparators described in this section?
What approaches might be used to identify and justify a comparator
such as unaided clinical judgment, delayed specialist review, or no
intervention?

Position Support

Recommendation to In behavioral health, the realistic alternative for many patients is a
FDA waitlist or no treatment. We recommend that benefit-risk
determinations account for that counterfactual, consistent with FDA's
guidance on uncertainty in benefit-risk determinations.

Citations Discussion paper, Section V.E, p. 18; FDA guidance, Consideration of
Uncertainty in Making Benefit-Risk Determinations (cited at p. 19, n.
21).

Question 16: Independent third parties

FDA question Should independent third parties be involved in some or all aspects of
a competency-based approach, including device benchmarking and
clinical confirmation? If so, in what ways might qualified, independent
third-party participation contribute to this approach, and what
qualifications and independence criteria should apply? Are there
aspects of the assessment for which third-party involvement would be
impractical or inadvisable? What safeguards or program design
 features would be critical to prevent third-party participation from
limiting competition or preventing innovation?

Position Support with modifications

Recommendation to Independence requirements should be proportionate to sponsor size
FDA and device risk, particularly where structurally independent human
adjudication is needed at scale.
For LLM adjudicators, independence should be demonstrated
empirically. That means validation against independent clinician
consensus, plus evidence that judge errors are not correlated with
device errors. Use of a different model family is a helpful but
incomplete safeguard. This applies to postmarket monitoring as well
as premarket adjudication.

Citations Discussion paper, Section V.B.2, p. 14; Section V.D.2, pp. 16–17.

Section VI: Postmarket monitoring

Question 18: Premarket uncertainty and postmarket reliance

FDA question CDRH is considering whether it may be appropriate to accept greater
premarket uncertainty regarding a GenAI-enabled device's benefit-risk
profile through greater reliance on postmarket monitoring. Under what
conditions might such an approach be appropriate, and what
characteristics of a monitoring program would need to be in place to
justify reduced premarket evidence? Are there device types or risk
profiles for which this approach would not be appropriate?

Position Conditionally support

Recommendation to Limbic supports greater reliance on postmarket evidence where the
FDA sponsor can change the device in controlled increments and measure
the effect of each change. We recommend CDRH recognize a model
in which each change (1) passes a pre-release gate proportionate to
the change, (2) is deployed to a limited population under monitoring,
and (3) is expanded or rolled back against prespecified criteria.
The following monitoring characteristics would justify reduced
premarket evidence: automated review of all interactions by validated
supervisory agents; timely clinician review of flagged sessions; rapid
rollback capability; and separate tracking of fast safety signals and
slower effectiveness outcomes.
We also ask CDRH to identify the legal mechanism that would make
postmarket commitments enforceable for each pathway (for example,
De Novo special controls, section 522 postmarket surveillance orders,
or PMA conditions of approval), so sponsors can plan accordingly.

Citations Discussion paper, Section VI, p. 19; FD&C Act section 522.

Question 19: Postmarket monitoring approaches

FDA question Please comment on the potential approaches to postmarket
performance evaluation, including periodic re-benchmarking,
sample-based clinician review, and performance degradation
 monitoring. What additional approaches should CDRH consider, and
how should the cadence and triggering events for reassessment be
determined?

Position Support with modifications

Recommendation to Limbic supports all three approaches and proposes two more.
FDA The first is targeted clinician review of device performance failures.
Supervisory agents route to clinicians the sessions where the device's
handling fell below prespecified criteria. This focuses clinician effort on
device errors and gives a direct measure of device performance.
The second is staged rollouts with change-level monitoring. Each
change is monitored separately, against prespecified rollback criteria
(see Q18).
Representative sampling should remain as an audit layer, specifically
a random sample of sessions that automated review did not flag, so
the miss rate can be measured. On cadence, we recommend
continuous automated review, with reassessment triggered by any
change and by monitoring thresholds.
We also ask CDRH to address the privacy constraints on reviewing
real patient interactions, particularly behavioral health transcripts.
These include the HIPAA minimum-necessary standard, state health
privacy laws, and data retention and deletion commitments.

Citations Discussion paper, Section VI.A, pp. 19–20; 45 CFR 164.502(b).

Question 20: Machine-based supervisory agents

FDA question Please comment on whether the proposed approaches to postmarket
monitoring can be facilitated by machine-based supervisory agents.
What considerations, including the evaluation and reliability of the
supervisory agent itself, should CDRH take into account for such an
approach?

Position Support with modifications

Recommendation to Supervisory agents make it practical to review every interaction. This
FDA matters for conversational devices, where safety-relevant events are
rare and may occur late in long conversations. We recommend that
CDRH treat a supervisory agent as a validated measurement
instrument, with six safeguards: (1) prespecified acceptance criteria
for each label, such as sensitivity and specificity against clinician
consensus, met before the agent may route sessions for review; (2)
continuous re-estimation of performance against ongoing clinician
labels, with a prespecified fallback to expanded human review if
performance falls below criteria; (3) a random audit of sessions the
agent did not flag, so its miss rate can be measured; (4) safeguards
against errors correlated with the device's own; (5) separation from
any automated judges or reward models used to optimize the device;
and (6) change control for the supervisor equivalent to that for the
device's safety controls.
We also recommend that CDRH recognize two modes of routing
sessions to clinicians. Situation-based routing sends a session to
clinicians when the patient's input indicates clinically significant
content, regardless of how the device performed. Performance-based
 routing sends sessions where the device's handling fell below
prespecified criteria. Situation-based routing provides broad coverage
early in deployment. Performance-based routing focuses clinician
effort on device failures and gives a direct measure of device
performance.
Sponsors should be able to step down from situation-based toward
performance-based routing against explicit criteria: demonstrated
supervisor performance that permits the step-down, and prespecified
triggers for stepping back up. Situation-based review should remain as
a floor for content that requires clinical determination regardless of
device performance, such as disclosures with mandatory-reporting or
duty-to-warn implications, and high-acuity risk.

Citations Discussion paper, Section VI.A, p. 20.

Question 22: Scaling re-benchmarking to modifications

FDA question CDRH envisions that a manufacturer's premarket competency-based
assessment might serve as a baseline against which post-deployment
modifications could be re-evaluated. How might the extent of
re-benchmarking or other evidence be scaled to the nature and
expected impact of a given modification? For example, are there
categories of change that might not significantly affect safety or
effectiveness of a GenAI-enabled device, or might be appropriately
managed within a sponsor's quality management system versus
requiring FDA premarket review and authorization before
implementation? Are there categories of changes that might be
appropriate for inclusion in a PCCP?

Position Support with modifications

Recommendation to Limbic recommends that review requirements scale with how much a
FDA change can expand what the device does. We propose four tiers.
Tier 1: QMS, with targeted regression testing. Bug fixes; latency
and infrastructure changes with no content change; and safety-layer
changes that only restrict behavior. Restrictive changes still need
targeted re-benchmarking for over-refusal and over-escalation,
because tightening a guardrail can increase both, and CDRH itself
treats those as failures (S.1, S.2).
Tier 2: QMS, with full automated re-benchmarking and a staged
rollout. Changes to prompts and therapeutic content within the
validated envelope.
Tier 3: PCCP, with prespecified re-benchmarking, staged rollout
with rollback criteria, and reporting to FDA. New or modified
therapeutic modules; changes to orchestration or routing logic;
foundation model version changes; and changes to monitoring agents.
Tier 4: New submission. A new intended use or population; new
clinical functions; any expansion of the actions the device can take;
and removal or loosening of a safety control.

Citations Discussion paper, Section VI.C, pp. 20–21; FDA guidance, Marketing
Submission Recommendations for a Predetermined Change Control
Plan for AI-Enabled Device Software Functions.

Question 23: Adapting PCCPs
FDA question How might PCCP concepts or other change-control approaches be
adapted for GenAI-enabled devices when the nature or scope of future
modifications cannot be fully prespecified?

Position Strongly support

Recommendation to Limbic recommends that a PCCP be permitted to prespecify the
FDA verification gate that a modification must pass. This would allow a
PCCP to cover modifications whose specific nature cannot be
described in advance. Under this approach, any modification within
the authorized intended use could be implemented if the device
passes both prespecified automated re-benchmarking and a staged
rollout with prespecified rollback criteria. The rollout step is needed
because simulation alone will not surface every failure mode (see
Q13).
Where devices are improved through automated optimization
techniques, each optimized version should be treated as a discrete
modification that passes the same gate before deployment. This is the
practical form of the continuous, iterative process described in our
overarching comments: frequent improvement, with every iteration
measured at the system level.

Citations Discussion paper, Section VI.C, pp. 20–21.

Question 24: Third-party foundation model changes

FDA question For devices built on third-party foundation models, changes to the
underlying model may be initiated by the third-party model developer
rather than the device manufacturer. How can a manufacturer detect,
evaluate, and respond to such changes in a timely manner? What
mechanisms, for example, contractual, technical, or through a PCCP,
could provide reasonable assurance that third-party developer-initiated
changes do not compromise the safety or effectiveness of the device?

Position Support with modifications

Recommendation to Limbic recommends that CDRH recognize a combination of controls
FDA as reasonable assurance: model version pinning, contractual minimum
deprecation-notice periods, and a PCCP modification protocol that
permits a model version change only after it passes prespecified
re-benchmarking.
A model version change should go through the same gate as any
other change, meaning re-benchmarking plus a staged rollout. As
models improve, upgrading is in patients' interest, at a pace that
per-version FDA review could not keep up with. The gate is what
makes routine upgrades feasible.
Vendors can also change safety-relevant behavior, such as content
filtering, without changing the version number. Production monitoring
should be designed to detect such changes, and contractual
commitments should cover notification of safety-behavior changes as
well as deprecations.

Citations Discussion paper, Section VI.C, p. 21.

Section VII: Other topics
Question 25: Foundation Model MAFs

FDA question Would voluntary Foundation Model MAFs be practical for foundation
model developers to provide and to keep current, and what content
might they include? Given that participation would be voluntary and
that model developers may have limited incentive to disclose
safety-relevant information, what would make such a program
sufficiently useful for premarket review? Are there alternative
mechanisms CDRH should consider for obtaining information about
underlying foundation models?

Position Support with modifications

Recommendation to Limbic supports the concept. Because participation would be
FDA voluntary, MAFs should be one of several acceptable mechanisms.
We ask CDRH to confirm that the device system remains the unit of
evaluation (Section V.A), and that system-level evidence is sufficient
where no MAF exists. That evidence would include version pinning,
re-benchmarking on model change, and contractual notification
commitments. Otherwise, MAFs risk becoming a de facto expectation.
Sponsors should not be expected to obtain and submit information
that a model developer will share only with FDA directly. If that
happened, a sponsor's ability to adopt a better model would depend
on a third party's disclosure decisions.
The most useful MAF content is information only the developer holds:
training-data overlap with public benchmarks (contamination);
change-notification commitments, including for safety-behavior
changes; and deprecation timelines. Static per-version model cards
will be outdated within months.

Citations Discussion paper, Section VII.A, p. 22 and n. 24.

Question 26: Agentic systems

FDA question Are there additional considerations that inform the premarket and
postmarket evaluation of agentic GenAI-enabled devices, beyond
those applicable to non-agentic GenAI-enabled devices? How should
the elevated risk associated with autonomous multi-step action, tool
use, and reduced opportunity for human review be reflected in
acceptance criteria and oversight?

Position Support with modifications

Recommendation to Limbic recommends that risk for agentic devices scale with the
FDA device's action envelope: the external actions the device can take,
their reversibility, and the controls gating them.
Devices composed of multiple cooperating models should not be
treated as higher risk on that basis alone if their external actions are
unchanged. Acceptance criteria should require deterministic gating
and human checkpoints proportionate to the consequences of
irreversible actions.

Citations Discussion paper, Section VII.B, pp. 22–23; Appendix A, A.1, p. 26.

Part 2. Additional issues not raised in the discussion questions
Issue 1: Reconciliation with the 2026 CDS and General Wellness guidance

Issue The paper indicates that CDRH is considering treating action-directing
functions as higher risk (p. 7), but does not discuss the January 2026
CDS guidance's enforcement discretion for a single clinically
appropriate recommendation, or the revised General Wellness
guidance. Conversational products can move between wellness,
non-device, and device functions within one session.

Recommendation to We recommend that CDRH explain how the risk framework in Section
FDA IV relates to the January 2026 Clinical Decision Support Software
guidance and the revised General Wellness guidance. In particular,
CDRH should clarify (1) whether an HCP-facing function that provides
a single clinically appropriate recommendation, and that would fall
within the enforcement discretion described in the CDS guidance,
would be treated differently because it is GenAI-enabled; and (2) how
the Multiple Function Device Products guidance applies when one
conversation moves between general wellness content, non-device
functions, and device functions.
For conversational products, we recommend that CDRH accept
function boundaries defined through the manufacturer's design
controls, such as routing and gating that determine which component
responds. Trajectory testing under benchmarking element S.2 would
confirm that those boundaries hold. This would allow sponsors to
apply existing policies consistently across a session.

Citations Discussion paper, Section III, pp. 4–5, and Section IV, p. 7 and n. 13.
FDA, Clinical Decision Support Software (January 2026). FDA,
General Wellness: Policy for Low Risk Devices (January 2026). FDA,
Multiple Function Device Products: Policy and Considerations.

Issue 2: Legal authority for proposed mechanisms

Issue The paper does not address authority for third-party certificates,
enforceable postmarket monitoring for 510(k) and De Novo devices, or
re-benchmarking triggered by changes outside the sponsor's control.

Recommendation to We recommend that future guidance identify, for each mechanism
FDA CDRH advances, the statutory authority it relies on and the pathways
to which it applies. Several relevant authorities already exist: section
515C for predetermined change control plans; section 522 for
postmarket surveillance orders; special controls under section
513(a)(1)(B) for De Novo and 510(k) devices; section 523 for
accredited third-party review of 510(k) submissions; and the
Accreditation Scheme for Conformity Assessment program under
section 514(d).
Where a mechanism would require new authority, such as binding
third-party certificates or postmarket monitoring conditions on cleared
devices, we ask CDRH to state that clearly. Sponsors can then
distinguish recommendations from requirements when planning
submissions.

Citations Discussion paper, p. 1. FD&C Act sections 513(a)(1)(B), 514(d), 515C,
522, and 523.

Issue 4: Interaction with existing device frameworks
Issue The paper does not discuss how the competency-based approach
relates to section 524B cybersecurity requirements (despite
prompt-injection testing in S.2 and A.1), human factors engineering, or
software lifecycle standards such as IEC 62304.

Recommendation to We recommend that CDRH confirm that evidence generated under
FDA existing frameworks can be cross-referenced in a competency-based
assessment, so sponsors avoid duplicative testing.
Adversarial and prompt-injection testing under S.2 and A.1 could also
support the cybersecurity documentation required under section 524B.
Summative usability testing conducted under FDA's human factors
guidance and IEC 62366-1 could support element E.4, including user
comprehension and automation bias.
Software life cycle documentation under IEC 62304, maintained within
the quality management system required by the Quality Management
System Regulation, could support the change control that
re-benchmarking depends on.
We also ask CDRH to align terminology with its guidance on life cycle
management for AI-enabled device software functions.

Citations FD&C Act section 524B. FDA, Cybersecurity in Medical Devices:
Quality System Considerations and Content of Premarket
Submissions. FDA, Applying Human Factors and Usability
Engineering to Medical Devices (2016). IEC 62366-1. IEC 62304. 21
CFR Part 820. FDA, Artificial Intelligence-Enabled Device Software
Functions: Lifecycle Management and Marketing Submission
Recommendations. Discussion paper, Appendix A, pp. 24–26.

Issue 5: International harmonization

Issue The paper does not reference IMDRF work, the EU AI Act, or MHRA
approaches. Divergent evidence frameworks increase burden for
sponsors operating in multiple jurisdictions.

Recommendation to We recommend that CDRH align the competency-based approach
FDA with the IMDRF Good Machine Learning Practice guiding principles,
finalized in January 2025, and with IMDRF's ongoing work on AI life
cycle management, using IMDRF terminology where possible.
IMDRF's 2026 draft technical framework on AI life cycle management
overlaps directly with the paper's benchmarking and postmarket
proposals.
We also recommend that CDRH accept benchmarking and clinical
evidence generated for other regulators where the methods meet
FDA's prespecified criteria, and that CDRH pursue alignment with
MHRA and EU authorities on change-control mechanisms equivalent
to PCCPs, building on the joint FDA, Health Canada, and MHRA
guiding principles. Aligned evidence expectations would reduce
duplicate testing for sponsors operating in multiple jurisdictions,
including those subject to the EU AI Act's requirements for high-risk AI
systems that are medical devices.
Citations: IMDRF/AIML WG/N88 FINAL:2025, Good Machine Learning
Practice for Medical Device Development: Guiding Principles (January
2025). IMDRF/AIML WG/N93 DRAFT, Technical Framework for
Artificial Intelligence Life Cycle Management (April 2026). FDA, Health
Canada, and MHRA, Predetermined Change Control Plans for
 Machine Learning-Enabled Medical Devices: Guiding Principles.
Regulation (EU) 2024/1689 (AI Act).

Issue 6: Transition for devices in FDA pilot programs

Issue The paper does not address how evidence generated by devices
already operating under FDA pilot programs or enforcement discretion
would be credited under a future framework.

Recommendation to We recommend that CDRH provide a transition approach for devices
FDA that have generated real-world evidence under FDA pilot programs or
enforcement discretion policies. Where that evidence was collected
under prespecified protocols, it should be eligible to support clinical
confirmation and to establish the baseline for postmarket
re-benchmarking. Sponsors should not have to repeat evaluations that
already meet the framework's criteria.
We also ask CDRH to give advance notice and a reasonable transition
period before applying new evidentiary expectations to devices
already in development or under review.
Citations: Discussion paper, Section V.C, pp. 14–16, and Section VI.C,
pp. 20–21.