SAS Institute Inc. (Campbell Buhrow)

Please see the attached industry feedback.

Title Page

Considerations for the Regulation
of Generative AI-Enabled Medical
Devices (FDA-2026-N-7874)


Federal Drug Administration
(FDA) Center for Devices and
Radiological Health (CDRH),


19 October 2026

Submitted by:
Campbell Buhrow
Account Executive
703-310-5722
Campbell.buhrow@sas.com

This proposal is the confidential and proprietary property of
SAS Institute Inc. It may contain approaches, techniques,
and other information proprietary to SAS, and shall not be
disclosed in whole or in part to third parties without the
prior written consent of SAS.

Copyright 2026 SAS Institute Inc. All Rights Reserved
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

Contents
Title Page ........................................................................................................ 1
Contents .......................................................................................................... i
Executive Summary ....................................................................................... 1
Discussion Questions ................................................................................... 2
Section IV: Considerations for the Assessment of Risk for GenAI-Enabled
Devices ....................................................................................................... 2
1. 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...................................2
2. 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? ......................................................................................3
3. 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? ..............................................................4
4. 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? 5
5. 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? ........................................................................................................5
6. 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? ........6
Section V: A Competency-Based Approach for Premarket Evaluation of
GenAI-Enabled Devices ............................................................................. 7
7. 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? ......................................................................................................7

FDA-2026-N-7874 Discussion Paper and Request for Feedback
Page i
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

8. 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? ...................................................................7
9. 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? .............8
10. 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? .......................................................................................................8
11. CDRH is considering that clinical confirmation for a GenAI-enabled device might
not require a prospective clinical study in every case and has described above a range of
approaches of increasing rigor and patient exposure. How might a sponsor select and
justify a confirmation approach tailored to a device’s intended use and proportionate to
the device’s risk profile? Are there device types or risk profiles for which one or more of
these approaches would be insufficient or inappropriate? How might the anticipated
distribution of real-world inputs be taken into consideration? Are there other methods of
clinical confirmation that might help inform the evaluation of GenAI-enabled devices?......9
12. How should sponsors achieve statistically meaningful performance measurement
for GenAI-enabled devices? Where synthetically generated inputs supplement real
patient data, how should sponsors account for differences between the synthetic and
real-world distributions when estimating performance, and under what conditions, if any,
is it appropriate to combine benchmarking evidence and clinical confirmation evidence to
support a single performance estimate? ............................................................................10
13. 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? ........................................................................................................................10
14. 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?....................................................................................................................11
15. 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?..................................................................................................................11
16. 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? .....................12
17. CDRH is exploring whether a competency-based approach could be applied to
devices that incorporate a variety of underlying model architectures, such as multimodal

FDA-2026-N-7874 Discussion Paper and Request for Feedback
Page ii
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

vision-language models and generative or predictive world models. Are there aspects of
the described approach that might be ineffective or inapplicable to these devices? What
additional or distinct considerations might need to be addressed? ...................................12
Section VI: Postmarket Monitoring for GenAI-Enabled Devices ............... 13
18. 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? ...........................................................................13
19. 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? .......................................................................................................................14
20. 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? ..........................................................................................14
21. What roles might clinicians, healthcare institutions, professional societies,
standards-setting bodies, and other stakeholders appropriately play in postmarket
monitoring, and how can these roles be structured without diffusing manufacturer
accountability? ...................................................................................................................15
22. CDRH envisions that a manufacturer’s premarket competency-based assessment
might serve as a baseline against which post-deployment modifications could be reevaluated. 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? ......15
23. 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? ......................................................................................................................16
24. 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? ..........................................16
Section VII: Other Topics .......................................................................... 17
25. 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? .........................17
26. 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 multistep action, tool use, and reduced opportunity for human review be reflected in
acceptance criteria and oversight? ....................................................................................17
Back Cover ................................................................................................... 18

FDA-2026-N-7874 Discussion Paper and Request for Feedback
Page iii
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

Executive Summary
FDA’s request for feedback on generative AI-enabled medical devices comes at an important
point in digital health. These technologies can expand access to clinical information, support
responsive care, and improve complex decisions. SAS supports CDRH’s effort to establish a
nimble, risk-based, and risk-based, and least burdensome approach that enables responsible
innovation while assuring safety and effectiveness across the total product life cycle.

Evidence requirements and oversight should be proportionate to a device’s activity and the
consequences of an incorrect output or analysis. FDA’s two-axis framework is a strong
foundation, but assessment should also consider reversibility, time pressure, user expertise,
traceability, safeguards, and error correction. Directiveness should be treated as a continuum,
with risk evaluated across realistic interactions, in context of human impact, complete
workflows, and the full configured device, not only the underlying model.

SAS supports a competency-based premarket approach that combines rigorous
benchmarking with clinical confirmation tailored to intended use and risk. Testing should use
prespecified methods, representative and well governed data, and meaningful acceptance
criteria, and should include uncertainty estimates, reproducibility checks, and subgroup
analysis as well as bias and fairness assessments. Clinical confirmation should establish
performance in representative settings with evidence scaled to the consequences of
decisions. Synthetic data can strengthen stress testing and rare-event evaluation but should
supplement, not replace, independently governed real-world evidence. Synthetic data should
also be screened for statistical fidelity to source data before it is used in clinical assessment.

Because GenAI-enabled devices may change after deployment, premarket evidence must
connect to continuous postmarket oversight. Manufacturers should maintain versioned
baselines, monitor for drift, conduct risk-based re-benchmarking, and define thresholds for
investigation and regulatory reporting. Predetermined Change Control Plans should bind
permitted changes and specify safety controls, testing, approvals, and monitoring. The same
discipline should govern third-party models and agentic systems through version control, drift
testing, human approval, traceability, and fail-safe controls. Across all models, it is essential
to have a predetermined emergency shut off as well as an understanding when a model
should be sunset or retired.

SAS has more than four decades of experience supporting the FDA and regulated life
sciences industries, including analytics for clinical trial evidence and regulatory review. That
experience informs a deliberate approach: apply deterministic and probabilistic methods
appropriately, be cautious and intentional with non-deterministic uses, preserve human
accountability, and build governance into the analytical environment. The SAS Platform
supports this approach natively. It provides governed model development, deployment, and
oversight across SAS, open-source, and third-party models. These capabilities help the FDA
and GenAI model providers keep evidence reproducible, decisions auditable, and patient
safety central.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

Discussion Questions
Section IV: Considerations for the Assessment
of Risk for GenAI-Enabled Devices
1. 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 GenAIenabled 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.
The two-axis framework is a sound foundation, and we recommend that it be treated as the
core of a multidimensional assessment rather than a complete risk determination. FDA’s
drug-development credibility framework similarly combines model influence with decision
consequence, while emphasizing context of use and risk-proportionate evidence. For
Generative AI-enabled devices, we recommend that CDRH require documented modifiers for
reversibility, time to detect and correct error, availability and independence of downstream
safeguards, user expertise, source traceability, output uncertainty, and scale of exposure.
These modifiers should not be averaged into a single opaque score. Instead, we recommend
that they be recorded as structured factors that can raise the evidence tier or impose use
restrictions. For example, an incorrect draft discharge summary that is reviewed before
release is materially different from the same output delivered directly to a patient without
review. The framework should evaluate the complete decision system, including human-AI
interaction, not only the model output. An integrated SAS Platform can enable this approach
through a governed use-case inventory, configurable multidimensional risk assessments,
linked evidence and data lineage, workflow-based review and approval, versioned controls,
and auditable monitoring of residual risk across the full decision system.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

2. 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?
Directiveness is best assessed as a continuum based on how an output is likely to influence
user action within its intended context, including consideration of the intended audience.
Characteristics that may modify risk include imperative or urgent wording, specificity of a
recommended action, personalization to patient data, ranking or omission of alternatives,
default selections, presentation of confidence, timing, repetition across interactions, and
whether the system initiates or facilitates downstream tasks. Importantly, the same output
may carry different levels of perceived urgency or influence depending on whether it is
presented to a healthcare professional (HCP) or a patient. For example, technical or clinical
terminology that an HCP may interpret neutrally (such as an X-ray result with “abnormal mass
detected”) could be misunderstood or perceived as more definitive or alarming by a patient.
Accordingly, disclaimers alone should not outweigh the practical effect of the output, its
presentation, or the user interface.

CDRH could improve clarity and regulatory predictability by providing representative
examples and a structured framework that distinguishes functions such as retrieval,
summarization, recommendation, prioritization, approval, and execution while recognizing that
these functions may fall at different points along the directiveness continuum depending on
context. Manufacturers could assess the maximum foreseeable level of directiveness,
consider whether users can inspect primary sources, seek independent confirmation, or
reverse resulting actions, and conduct robust user-experience and human-factors testing with
intended user populations to identify potential misunderstandings or unintended
consequences before release. Testing should include representative users and scenarios to
evaluate not only what the GenAI system communicates, but how users are likely to interpret
and act upon that information.

The SAS Platform can support this type of risk-based assessment through configurable
decision flows, policy rules, role-based approvals, source-linked outputs, scenario testing,
traceability, and Natural Language Explanations that can make technical outputs more
understandable to different audiences. These capabilities can help manufacturers document
how informational outputs are generated and presented, assess how they may progress
toward consequential actions, and implement controls designed to prevent an informational
function from unintentionally crossing into an unauthorized action-directing role.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

3. CDRH seeks input on whether and when a GenAIenabled 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?
Patient-facing delivery of clinical information may present a different risk profile than delivery
of the same information to a healthcare professional, particularly when users lack the clinical
context, domain knowledge, time, or access needed to independently evaluate or challenge
an output. At the same time, patient capability should not be presumed to be low. Rather, the
relevant consideration should be whether the device and its GenAI-enabled functions are
designed to enable informed and safe use by the intended patient population.

Manufacturers should consider how the wording, specificity, personalization, and presentation
of clinical information may affect patient interpretation. Patients may reasonably interpret
generalized, population-based, or probabilistic information as personally applicable,
particularly when it concerns a condition, symptom, or outcome directly relevant to their
health. Appropriate safeguards may therefore include plain-language explanations, clear
descriptions of system limitations, visibility into sources and underlying data, calibrated
expressions of uncertainty, and a clear distinction between educational information and
diagnosis or treatment recommendations. Manufacturers should also consider accessible
design, localization, comprehension testing, escalation to qualified healthcare professionals,
and safe-failure approaches when supporting evidence is insufficient. These measures can
help mitigate unnecessary alarm, false reassurance, inappropriate self-management, or
overreliance without unnecessarily restricting patients' access to or ability to understand their
own clinical information.

For higher-consequence outputs, manufacturers should not rely solely on generic disclaimers
or the nominal availability of human support. Robust user-experience and human-factors
testing should evaluate how intended users understand and respond to outputs in realistic
settings, including users with varying levels of health literacy, language proficiency,
accessibility needs, and experience with clinical information, as well as users encountering
stressful or urgent situations. Such testing can identify where additional context or safeguards
are warranted while avoiding paternalistic restrictions that are not tied to demonstrated risk.

Risk assessments should also consider foreseeable use outside the intended user population
or use case. GenAI-generated outputs may be accessed, shared, or relied upon by patients
or healthcare professionals in ways that extend beyond the device's intended use.
Manufacturers should evaluate whether foreseeable off-label use could increase the potential
for misunderstanding, overreliance, or inappropriate action, and whether audience-specific
presentation, contextual information, or additional controls could mitigate those risks.

The SAS Platform can support these safeguards through audience-specific presentation
rules, localization, provenance and recency information, confidence thresholds, accessibility
and comprehension analytics, governed escalation pathways, and monitoring of patient-facing

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

interactions for unsupported, misunderstood, or potentially high-risk outputs. Together, these
capabilities can support a risk-based approach that recognizes differences between patient
and HCP use while preserving the potential benefits of GenAI for patient empowerment,
engagement, and access to clinical information.

4. 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?
The distinction between generalist and specialist physicians should be based on task-specific
competence, not professional title alone. Risk increases when a device presents specialistlevel conclusions to an HCP who may be less likely to recognize omissions, atypical
presentations, or inappropriate certainty, especially when the output directs action. Devices
should clearly display the assumed audience, including whether information is intended for a
specialist with deep domain knowledge or a generalist, enabling HCPs to appropriately
interpret the output and recognize when specialist consultation may be needed. Safeguards
may include scope-aware routing, specialty-specific warnings, abstention outside validated
domains, evidence links, displays of uncertainty and alternative explanations, mandatory
review for defined high-risk patterns, and auditable transfer to a specialist. Evaluation should
include realistic generalist-to-specialist handoffs and assess whether specialist input is
appropriately identified. The SAS Platform can support this design through identity- and roleaware routing, competency metadata, specialty-specific policy rules, uncertainty-based
escalation, governed review queues, and audit trails that document consultation, override,
and handoff outcomes.

5. For multi-turn conversational GenAI-enabled
devices that may migrate from providing “nondirective” 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?
Risk should be assessed across bounded conversational trajectories, including foreseeable
escalation from general information to personalized recommendations or actions. The
intended use should define permitted intents, prohibited transitions, maximum actionability,
tool permissions, memory behavior, and conditions requiring abstention or handoff. Testing
should cover realistic multi-turn paths, adversarial prompts, conflicting facts, accumulating
personalization, and attempts to bypass guardrails. Because emergent behavior may not be
captured by single-turn benchmarks, manufacturers should evaluate conversation-level
outcomes and retain sufficient traces to reconstruct consequential exchanges. Manufacturers

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

should also be encouraged to log the sentiment of outputs of the system, allowing them to
better understand when a system transitions from informational to directive. This holistic
visibility will allow them to recognize patterns and increase the robustness of testing for the
continuously deployed systems. Model cards can also be used to document the intended use
and out of scope uses of a model. Audit logs of systems can be regularly compared to the
model cards to ensure no drift in scope has occurred. If any reasonably foreseeable trajectory
becomes action-directing, the device should be evaluated and controlled for that higher-risk
state unless the transition is demonstrably blocked. The SAS Platform can enable trajectorylevel testing and governance by versioning prompts, policies, memory settings, tools, and
retrieval sources. It can simulate multi-turn scenarios, enforcing permission boundaries and
retaining reconstructable conversation and action traces for review. It can also provide Model
Cards for each deployed and developed model, allowing for scope documentation.

6. 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?
Under-escalation and over-escalation should be evaluated separately because their harms,
frequencies, and tolerability differ by clinical context. Decisions regarding care escalation are
also not purely statistical or medical in nature but may depend on the values, principles, and
risk tolerances of the manufacturer and healthcare organizations deploying the technology. In
clinical contexts where patient well-being, beneficence, and the principle of “do no harm” are
central, manufacturers should explicitly evaluate the consequences of both directions of error.
Under-escalation may result in delayed or missed treatment as well as emotional,
psychological, and quality-of-life impacts, while over-escalation may result in patient anxiety,
unnecessary testing or treatment, resource utilization, alarm fatigue, or disruption of care.
Sponsors should predefine clinically grounded cost matrices, report sensitivity and specificity
with confidence intervals, and establish directional-error limits, subgroup results, calibration,
time-to-escalation, override behavior, and explicit fallback rules. Thresholds should be
selected with domain experts and tested prospectively or in realistic simulation under
expected prevalence and workflow conditions, as aggregate accuracy or a single weighted
score can conceal unacceptable performance in either direction. The relative weighting of
these risks may differ across clinical contexts, but the rationale for those trade-offs should be
transparent, documented, and aligned with clearly articulated clinical and ethical objectives.
The SAS Platform can support separate directional-error analytics, calibrated threshold
optimization, subgroup and fairness assessment, simulation at expected prevalence,
operational alert testing, and governed approval of context-specific thresholds and fallback
actions.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

Section V: A Competency-Based Approach for
Premarket Evaluation of GenAI-Enabled Devices
7. 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?
Yes, a competency-based approach is useful if benchmarking and clinical confirmation are
linked to a clearly defined context of use and are scaled the model influence and the
consequence of the decision. Benchmarking can establish controlled evidence for knowledge,
reasoning, communication, robustness, and safety behavior, while clinical confirmation tests
performance in the actual human, workflow, and data environment. Neither tier should be
treated as sufficient on its own for consequential uses. The framework should require
prespecified acceptance criteria, independent test data, uncertainty estimates, subgroup
analysis, reproducibility, and documented deviations. For systems that include retrieval, rules,
tools, or human review, the subject of evaluation should be the complete configured device,
not only the underlying foundation model. The SAS Platform can connect benchmark
definitions, controlled test data, statistical results, clinical-confirmation evidence, system
versions, deviations, and approvals in a traceable evidence workflow, supporting reproducible
execution across SAS, R, and Python.

8. How might the two-axis risk framework described
in Section IV be considered within the competencybased approach described above to help determine
the level of evidence needed for a premarket
submission?
The risk framework should determine the breadth, independence, realism, and statistical rigor
of the evidence required, while also considering the degree of uncertainty and variability
associated with a specific non-deterministic model or system. For GenAI-enabled functions,
the same task may yield materially different outputs depending on user inputs, conversational
context, available information, or workflow conditions. Greater model influence, decision
consequence, variability, or uncertainty should therefore trigger stronger data independence,
larger and more representative samples, tighter confidence bounds, more stringent
directional-error limits, human-factors testing, external or independent review, and more
extensive clinical confirmation. Lower-risk uses may justify narrower testing if safeguards
materially reduce reliance on the output. CDRH should avoid a purely numerical risk score
that obscures judgment. Sponsors should provide separate rationales for model influence,
consequence, uncertainty and variability, and other modifiers, followed by a transparent
mapping to evidence tiers and residual controls. The SAS Platform can implement this
mapping through configurable risk tiers that trigger required evidence, independent-review
steps, test templates, statistical acceptance criteria, and release gates, while preserving the
rationale and approvals rather than relying on an opaque composite score.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

9. 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?
The proposed categories are broadly appropriate, but CDRH should add source faithfulness,
uncertainty calibration, boundary adherence, tool-use safety, human-AI team performance,
robustness to change, and other characteristics unique to generative AI systems as explicit
elements. While benchmarks of clinical knowledge can help assess whether a system can
perform its intended function, additional measures such as variability, temperature-based
assessments, toxicity, refusal behavior, and faithfulness can help evaluate whether the
system behaves appropriately when presented with out-of-scope, harmful, or vague requests,
declines unsafe requests, remains grounded in appropriate evidence, and avoids harmful
outputs. Each construct should have a defined claim, representative tasks, scoring method,
acceptance threshold, and evidence of inter-rater reliability where human judgment is used.
Clinical knowledge should not be inferred from fluent communication, and communication
quality should not compensate for unsafe reasoning. Useful external foundations include the
NIST AI RMF and its Generative AI Profile, relevant ISO/IEC AI risk and lifecycle standards,
human-factors standards, and domain-specific clinical performance standards, applied where
their scope and validation evidence fit the device context. The SAS Platform can support a
reusable evaluation framework with metric libraries, custom scoring, human adjudication
workflows, inter-rater reliability analysis, source-faithfulness checks, subgroup analytics, and
a controlled link between each result and the exact model and system configuration tested.

10. 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?
Benchmark validity should be demonstrated by a documented chain from the intended-use
claim to the tasks, inputs, outputs, scoring rules, and clinically meaningful outcomes.
Sponsors should characterize overlap with development data, contamination risk, saturation,
prevalence, subgroup coverage, temporal and site variation, and the relationship between
benchmark performance and real-world failures. Public benchmarks are useful for
comparability but should not be the sole basis for consequential uses. Sponsor-developed
benchmarks can address proprietary workflows and rare hazards, provided test custody of the

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

test set is independent, cases are not reused for tuning, scoring is locked, and results are
confirmed on external or prospectively collected data. Construct validity also requires
evidence that higher benchmark scores predict safer or more effective behavior in the
intended setting. The SAS Platform can enable governed benchmark catalogs, dataset
lineage and overlap analysis, independent test-data access, locked scoring pipelines,
reproducible statistical evaluation, and a link from benchmark results to external or
prospective confirmation evidence.

11. CDRH is considering that clinical confirmation for
a GenAI-enabled device might not require a
prospective clinical study in every case and has
described above a range of approaches of increasing
rigor and patient exposure. How might a sponsor
select and justify a confirmation approach tailored to
a device’s intended use and proportionate to the
device’s risk profile? Are there device types or risk
profiles for which one or more of these approaches
would be insufficient or inappropriate? How might the
anticipated distribution of real-world inputs be taken
into consideration? Are there other methods of
clinical confirmation that might help inform the
evaluation of GenAI-enabled devices?
Sponsors should select the method clinical confirmation based on the intended use, model
influence, decision consequence, novelty, uncertainty, and the adequacy of available realworld evidence. Retrospective simulation may be appropriate for bounded, low-influence
functions with reliable reference outcomes. Silent prospective studies can test workflow
performance without affecting care. Supervised deployment may be needed when human
interaction materially changes outcomes; prospective interventional evidence is appropriate
when the device directly determines high-consequence decisions or existing evidence cannot
resolve uncertainty. The anticipated input distribution should be characterized across sites,
time, prevalence, devices, languages, and subgroups, with stress testing of credible
distribution shifts. For any approach, sponsors should predefine acceptance criteria and
explain why the chosen design measures the actual clinical claim. The SAS Platform can
support retrospective and prospective study analytics, real-world-data preparation and
lineage, simulation, site and subgroup comparisons, acceptance-criteria testing, and
governed integration of clinical-confirmation results with the device’s risk and intended-use
record.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

12. How should sponsors achieve statistically
meaningful performance measurement for GenAIenabled devices? Where synthetically generated
inputs supplement real patient data, how should
sponsors account for differences between the
synthetic and real-world distributions when estimating
performance, and under what conditions, if any, is it
appropriate to combine benchmarking evidence and
clinical confirmation evidence to support a single
performance estimate?
Statistically meaningful measurement requires prespecified endpoints, clinically relevant error
definitions, representative sampling, confidence intervals, subgroup analysis, and methods
that account for repeated measures and clustered data. Synthetic inputs should be analyzed
separately from real patient data unless a justified transportability analysis shows that the
synthetic distribution, labeling process, and error structure support pooling. Sponsors should
quantify distributional differences and report sensitivity analyses with and without synthetic
cases. Synthetic data are most defensible for stress testing, rare edge cases, privacypreserving development, and controlled perturbations. They should not inflate the effective
sample size for real-world claims. Benchmarking and clinical confirmation may inform a
combined conclusion, but ordinarily should retain separate estimates unless the designs,
populations, outcomes, and weighting are prospectively justified. The SAS Platform can
enable complex-sample and clustered-data analysis, confidence intervals, subgroup
estimates, distribution-comparison diagnostics, sensitivity analyses, privacy-preserving
synthetic-data workflows, and separate evidence tracks for benchmark, synthetic, and clinicalconfirmation results.

13. For which clinical domains, device functions, or
subpopulations is synthetic data particularly wellsuited, 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?
Synthetic data are particularly useful for controlled perturbations, privacy-preserving
development, workflow testing, rare but well-characterized hazards, and manufacturing or
imaging simulations with credible mechanistic constraints. They are less suitable where
disease heterogeneity, social context, clinician behavior, missingness, treatment selection, or
underrepresented populations are poorly modeled. Safeguards should include independent
real-world confirmation, disclosure of generator lineage, and tests for memorization and
privacy leakage. They should also include comparisons of marginal, joint, temporal, and
subgroup distributions. Evaluation methods should not share the same architecture or training
data as the device. Synthetic cases should not be used to claim the absence of bias in a
subgroup that is inadequately represented in real data. The SAS Platform can support
controlled synthetic-data generation, privacy-risk and similarity assessment, marginal and

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

multivariate distribution comparisons, temporal and subgroup diagnostics, lineage of
generation methods and seed data, and side-by-side validation against independently
governed real-world datasets.

14. 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?
Comparators and acceptance criteria should follow the clinical claim and intended workflow.
For open-ended outputs, a structured rubric can assess factual correctness, completeness,
harmful omissions, source support, uncertainty, communication, and actionability, with critical
errors weighted separately from stylistic differences. A clinician panel should be qualified for
the task, independent, blinded where feasible, and supported by adjudication and inter-rater
reliability analysis. The target should be a justified standard of care or other clinically
meaningful performance level, not automatically the performance of the median clinician.
Specialists should review specialist content, while intended users should participate in
human-factors evaluation. Human-AI team performance should be the primary measure
whenever the labeled use requires human review. Device-only testing remains necessary to
characterize intrinsic failure modes. The SAS Platform can support rubric-based review, blind
case assignment, clinician-panel adjudication, inter-rater reliability analysis, comparator and
non-inferiority testing, and separate measurement of device-only, clinician-only, and humanAI team performance.

15. 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?
Evaluation against the likely counterfactual can provide decision-relevant evidence. Sponsors
should define the actual alternative pathway, such as unaided judgment, delayed specialist
review, an existing rules engine, or no intervention using workflow observation, historical data,
pragmatic trials, or carefully designed simulation. The comparator should reflect current
practice at representative sites and account for delays, access constraints, and treatment
availability. Comparative measures may include diagnostic or triage accuracy, time to
appropriate action, unnecessary escalation, missed harms, workload, and patient outcomes.
Historical controls require careful adjustment for temporal and selection bias. A weak local

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

baseline should not justify an unsafe device. The SAS Platform can support counterfactual
and comparative-effectiveness analyses, propensity-score and time-to-event methods,
workflow simulation, site-level variation assessment, and governed documentation of
comparator selection, assumptions, and sensitivity analyses.

16. 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?
Independent third parties can strengthen confidence in test design, dataset custody,
contamination assessment, cybersecurity testing, statistical analysis, and selected clinical
confirmation, particularly for high-risk or widely deployed devices. Independence criteria
should address financial interests, prior involvement of model development, data access,
conflicts, competence, and freedom to publish or communicate material findings. FDA should
avoid requiring a single proprietary benchmark owner or certification pathway. Sponsors must
remain accountable for the submission and quality system, and some evaluations may
appropriately remain internal when patient privacy, proprietary information, or rapid incident
response makes external participation impractical. Common protocols, transparent
qualification criteria, auditability, and multiple eligible assessors can reduce barriers to
competition. The SAS Platform can provide segregated workspaces, role-based access,
controlled test-data custody, standardized assessment workflows, reviewer qualification and
conflict records, immutable version histories, and auditable exchange of evidence with
multiple independent assessors.

17. CDRH is exploring whether a competency-based
approach could be applied to devices that incorporate
a variety of underlying model architectures, such as
multimodal vision-language models and generative or
predictive world models. Are there aspects of the
described approach that might be ineffective or
inapplicable to these devices? What additional or
distinct considerations might need to be addressed?
The approach can extend to multimodal and world models, but modality-specific and temporal
risks require additional evidence. Multimodal systems should be tested for cross-modal
conflict, missing or corrupted modalities, modality dominance, provenance, alignment, and
performance when one input source shifts. Vision-language systems require clinically
appropriate image-quality and localization assessment in addition to text evaluation.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

Predictive world models require validation of temporal consistency, state representation,
simulation fidelity, compounding error, and uncertainty over longer horizons. For all
architectures, CDRH should evaluate the complete configured device, including
preprocessing, retrieval, tools, prompts, memory, and downstream actions, and should
require architecture-appropriate stress tests rather than a single universal benchmark. The
SAS Platform can orchestrate multimodal-data preparation, provenance, feature and signalquality assessment, architecture-specific metrics, scenario simulation, uncertainty analysis,
and integrated testing of preprocessing, models, retrieval, policies, and downstream decision
flows. For devices that incorporate connected sensors, wearables, or other streaming inputs,
the SAS Platform can support real-time data ingestion, event-time alignment, temporal feature
generation, sensor-quality monitoring, and evaluation of model behavior when sensors drift,
disconnect, or produce conflicting observations.

Section VI: Postmarket Monitoring for GenAIEnabled Devices
18. 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?
Greater reliance on postmarket monitoring is appropriate only when residual uncertainty
concerns performance that can be detected promptly, before irreversible harm occurs, and
when the expected benefit of timely access justifies that uncertainty. Preconditions should
include a narrowly defined intended use, effective human oversight, reliable exposure and
outcome capture, validated monitoring, predefined thresholds, rapid containment and
rollback, transparent user communications, and enforceable update controls. Greater reliance
is generally inappropriate when a single error can cause immediate catastrophic harm,
outcomes are delayed, monitoring coverage is incomplete, or corrective action cannot protect
affected patients. Postmarket evidence should complement, not replace, a minimum
premarket showing of safety and efficacy. The SAS Platform supports governed postmarket
plans, streaming and batch performance surveillance, exposure and outcome reconciliation,
threshold-based alerts, case investigation, escalation, rollback of managed model versions,
and traceable linkage between field signals and corrective or preventive actions. Where the
device produces or consumes connected-device telemetry, the SAS Platform streaming
analytics supports continuous surveillance of sensor quality, operational state, patient or
equipment signals, and predefined safety events, with real-time escalation when approved
limits are exceeded.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

19. 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?
CDRH should expect manufacturers to combine periodic re-benchmarking, statistically
designed clinician review, real-world performance monitoring, safety-event surveillance,
complaint analysis, analysis of override and escalation trends, input-distribution monitoring,
and targeted evaluation after material changes. Cadence should be risk-based and
supplemented by event triggers such as performance-threshold breaches, shifts in patient
population or disease prevalence, new sites or languages, upstream model or retrieval
changes, cybersecurity events, unexpected use patterns, or evidence of subgroup harm.
Monitoring plans should specify denominator data sources, handling of delayed labels, alert
ownership, investigation timelines, and actions ranging from increased sampling to restriction,
rollback, or suspension. Periodic review remains necessary when event triggers are in place
because threshold-based alerting alone may miss gradual or interacting changes. The SAS
Platform can combine scheduled re-benchmarking, statistical sampling, drift and performance
monitoring, complaint and override analytics, event-triggered workflows, delayed-outcome
reconciliation, and dashboards that connect alerts to owners, investigations, decisions, and
remediation. For connected-device deployments, the SAS Platform can extend monitoring to
device-fleet telemetry, sensor drift, latency, connectivity loss, firmware or configuration
changes, and emerging event patterns, enabling reassessment triggers based on both model
performance and the condition of the physical data-generating environment.

20. 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?
Machine-based supervisory agents may assist with surveillance, sampling, anomaly
detection, policy checks, and routing, but they create an additional safety-critical component
that must itself be specified, validated, monitored, and governed. Evaluation should address
false reassurance, correlated failure with the primary model, prompt injection, including
through tool inputs, privilege boundaries, data access, alert fatigue, independence, and safe
behavior when evidence is missing. The supervisory agent should not be the sole detector of
failures that share its data, architecture, or vendor dependencies. Manufacturers should apply
a defense-in-depth approach that combines deterministic controls, independent statistical
monitoring, human review, and tested shutdown or rollback procedures. All consequential
supervisory actions should be attributable and reconstructable. The SAS Platform could
support supervisory workflows with deterministic policy checks, anomaly detection,
independent statistical controls, restricted tool permissions, human review queues, alertquality analytics, and end-to-end logging of observations, recommendations, actions, and
overrides. Streaming IoT analytics could provide an additional supervisory layer by comparing
model actions with contemporaneous sensor and equipment conditions, detecting anomalous
event sequences, and initiating governed alerts or safe-state workflows without making the
supervisory mechanism the sole safety control.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

21. What roles might clinicians, healthcare
institutions, professional societies, standards-setting
bodies, and other stakeholders appropriately play in
postmarket monitoring, and how can these roles be
structured without diffusing manufacturer
accountability?
Clinicians and institutions can contribute real-world outcomes, workflow observations,
complaints, and structured case reviews. Professional societies can define clinical constructs,
minimum datasets, and reference practices. Standards bodies can support interoperable
metrics and reporting formats, and patient groups can identify communication, access, and
harm concerns. These roles should be governed by data-quality requirements, conflict-ofinterest controls, timelines, escalation routes, and confidentiality protections. Manufacturer
accountability should remain explicit. External input does not transfer responsibility for
monitoring, investigation, corrective action, or regulatory reporting. A single accountable
owner at the manufacturer should reconcile signals across contributors and document the
disposition of each signal. The SAS Platform could provide governed data-ingestion and
quality controls, standardized reporting interfaces, stakeholder-specific workspaces, casereview and escalation workflows, confidentiality and access controls, and a consolidated
signal record with clear manufacturer ownership and disposition.

22. 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?
Re-evaluation should be scaled according to what changed, which hazards could be affected,
and how the change alters the authorized, intended use. Low-impact changes may include
documentation corrections or infrastructure changes shown not to affect outputs. These can
be managed under the quality system with regression evidence. Data-pipeline, prompt,
retrieval, dependency, calibration, interface, or user-interface changes generally require
targeted testing. Changes to intended use, autonomy, patient population, clinical logic,
foundation model, or safety-critical performance may require broad re-benchmarking, clinical
confirmation, and regulatory review. A PCCP is most suitable for foreseeable, bounded
modifications with defined methods, acceptance criteria, monitoring, and implementation
controls. The premarket baseline should be versioned so each modification can be compared
against the relevant authorized configuration. The SAS Platform can inventory all relevant
components, assess change impact against authorized baselines, trigger risk-scaled

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

regression suites and approvals, compare performance across versions, and preserve
release, monitoring, rollback, and regulatory-decision evidence. Change-impact assessment
could also include connected-device dependencies such as sensor calibration, firmware,
sampling frequency, edge-processing logic, communication protocols, and equipment
configuration because these changes can alter the model’s effective input distribution without
changing the Generative AI model itself.

23. How might PCCP concepts or other changecontrol approaches be adapted for GenAI-enabled
devices when the nature or scope of future
modifications cannot be fully prespecified?
A PCCP for generative AI should prespecify the envelope of change rather than every future
modification. The plan should define change categories, protected safety properties,
prohibited changes, affected components, data management and evaluation methods,
minimum test coverage, acceptance thresholds, human approvals, deployment controls,
monitoring, and rollback. Changes outside that envelope should require a new assessment
and, where applicable, a new marketing submission. Because emergent behavior can arise
from interactions among components, the plan should cover prompts, retrieval corpora, tools,
policies, memory, dependencies, and workflow configuration, not only model weights. Broad
statements such as “performance improvements” are insufficient without measurable
boundaries and evidence requirements. The SAS Platform can encode the approved change
envelope as governed policies, maintain lineage across models, prompts, data, retrieval,
tools, and rules, automatically route out-of-envelope changes for review, and require
documented tests, approvals, deployment controls, and rollback readiness before promotion.

24. 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?
Supplier agreements should provide notice of material changes, stable version identifiers,
access to change documentation, security and data-use terms, defined support periods, and
rights to retain prior versions and roll back to them. Technical controls should pin versions
where feasible, continuously record the endpoint and configuration used, run canary and
regression evaluations, detect behavioral and distribution shifts, and isolate deployment when
thresholds fail. A PCCP may cover bounded vendor-initiated changes only when the change
categories, evidence, acceptance criteria, and response are sufficiently specified. Silent
model substitution should be treated as a supplier-control failure. If the manufacturer cannot
obtain information needed to evaluate safety and effectiveness, the model should not be used
for that device function. The SAS Platform can maintain a governed third-party asset
inventory, capture endpoint and version metadata, run scheduled canary and regression
tests, monitor behavioral and input shifts, enforce deployment gates, and initiate isolation,
fallback, or rollback workflows when an upstream change breaches predefined thresholds.

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

Section VII: Other Topics
25. 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?
Voluntary Foundation Model MAFs could be useful if they provide current, reviewable
information that device manufacturers cannot independently generate. Core content should
include version identity, architecture and intended uses and known limitations. It should
describe training-data categories and governance, evaluation methods and subgroup results,
and cybersecurity and misuse testing. It should also cover change history and notification
practices, interfaces and tool-use constraints, data retention, incident information, and
reproducibility artifacts sufficient for downstream evaluation. Utility would increase through
standardized content, confidentiality protections, versioned updates, independent assurance,
and incentives such as reduced duplicative review. Alternatives include enforceable supplier
documentation requirements, confidential regulatory submissions cross-referenced by
multiple manufacturers, qualified third-party assessments, and standardized changenotification APIs. A master file should supplement, not replace, device-specific evidence for
the actual configured use. The SAS Platform can ingest and govern master-file metadata,
map supplier evidence to downstream use cases, control confidential access, track versions
and attestations, identify documentation gaps, and trigger reassessment when supplier
information, model versions, or use conditions change.

26. 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?
Agentic devices require additional evaluation because multi-step planning, memory, tool use,
and cross-system actions can compound error and reduce opportunities for review.
Acceptance criteria should cover task completion and clinical outcome as well as action
authorization, sequence validity, tool selection, source faithfulness, privilege containment,
reversibility, cumulative error, and safe interruption. Manufacturers should test against hostile
instructions, unavailable tools, conflicting evidence, long-horizon tasks, repeated actions, and
attempts to exceed scope. Autonomy should be risk-tiered. Read-only assistance, drafting,
recommendations, and consequential execution should have progressively stronger controls.
High-consequence actions should require authorization by a named person unless a specific,

FDA-2026-N-7874 Discussion Paper and Request for Feedback
 The SAS Platform: Advancing GenAI Innovation with Evidence, Governance, and Patient Safety

validated rationale supports otherwise. Complete audit-trail reconstruction, rate and resource
limits, least-privilege access, independent monitoring, fail-safe defaults, and rapid rollback
should be demonstrated for the entire agentic system. The SAS Platform can enable
governed agent registration, least-privilege tool access, policy-based action limits, human
approval checkpoints, scenario and adversarial testing, independent supervisory analytics,
complete action-chain logging, and automated containment or rollback when an agent
exceeds its approved scope. When an agent can interact with connected devices or physical
equipment, the SAS Platform can also support event-driven policies, command authorization,
operating-envelope checks, rate limits, independent telemetry monitoring, and safe-state
escalation, while keeping observation, recommendation, and physical actuation separately
Back Cover

To contact your local SAS office, please visit: sas.com/offices

SAS and all other SAS Institute Inc. product or service names are registered trademarks or
trademarks of SAS Institute Inc. in the USA and other countries. ® indicates USA registration. Other
brand and product names are trademarks of their respective companies.
Copyright 2026 SAS Institute Inc. All Rights Reserved.

FDA-2026-N-7874 Discussion Paper and Request for Feedback