i-GENTIC AI (Zahra Timsah)

Please see the attached comments submitted by Dr. Zahra Timsah, Founder and Chief Executive Officer of i-GENTIC AI, regarding FDA Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical Devices.

i-GENTIC AI | FDA-2026-N-7874

PUBLIC COMMENT

Considerations for the Regulation of
Generative AI-Enabled Medical Devices
Docket No. FDA-2026-N-7874 | Submitted September 23, 2026

Submitted by Dr. Zahra Timsah, PhD, MBA, MSc

Organization i-GENTIC AI

Role Founder and Chief Executive Officer

Location Houston, Texas

Executive recommendation
FDA should preserve the discussion paper's risk-proportionate and competency-based direction, but
extend the framework from periodic evaluation to continuous, evidence-producing governance across the
total product life cycle. Generative and agentic systems can change behavior between formal
assessments because of context, tool access, retrieval sources, configuration, third-party model updates,
or multi-step interactions. For higher-consequence functions, safety therefore cannot depend on
benchmark scores alone.

The regulatory objective should be not only to determine whether a device
performed acceptably in a test environment, but also to establish whether the
deployed system remains within an authorized safety envelope before, during, and
after each consequential action.

i-GENTIC AI recommends that CDRH define a minimum runtime governance evidence package for
higher-risk GenAI-enabled devices. That package should include machine-readable safety policies;
versioned provenance linking obligations, controls, tests, and deployed configurations; pre-action policy
checks; documented human-oversight checkpoints; immutable audit records; continuous performance and
drift monitoring; and a validated fail-safe mechanism capable of stopping or constraining unsafe action.

Perspective of the commenter
I submit these comments as Founder and CEO of i-GENTIC AI, a company focused on runtime governance
for regulated environments. My perspective reflects approximately seventeen years across medical
devices, biotechnology, healthcare, financial services, and AI governance, including engagement with
regulatory and public-policy stakeholders. These comments address governance architecture and
evidence principles. They do not disclose confidential customer information or proprietary implementation
details.

Public comment | Page 1
i-GENTIC AI | FDA-2026-N-7874

1. Risk assessment should measure controllability, not only activity
and consequence
Response to Questions 1-6. The proposed activity-by-consequence framework is a useful starting point,
but two functions with similar intended uses may present materially different real-world risk depending on
their autonomy, access, traceability, and capacity to be interrupted. CDRH should treat the following as
explicit risk modifiers:

• Autonomy and privilege: whether the system only generates information or can call tools, alter records,
initiate orders, control another device, or trigger downstream workflows.
• Reversibility and time to intervention: whether an action can be undone, whether harm could occur
before human review, and whether a safe rollback path exists.
• Traceability and provenance: whether an output can be linked to authoritative source material, the
applicable device configuration, the underlying model version, retrieved content, tool calls, and
governing policy.
• Detectability: whether an incorrect output or policy violation is likely to be recognized before reliance or
harm.
• Changeability: whether model, prompt, retrieval corpus, tool, workflow, or policy changes can alter
behavior without a clearly bounded release event.
• Downstream safeguards: whether independent controls, human checkpoints, permission limits, and
fail-safe mechanisms can prevent or contain harm.
Patient-facing versus HCP-facing status should not be a categorical proxy for risk. The more relevant
questions are whether the intended user can independently review the basis for an output, whether the
output is presented with unwarranted certainty, whether time permits verification, and whether the device
can directly cause or initiate a consequential action. A specialist-facing system may still be high risk when
it creates automation bias or acts faster than meaningful review can occur.

For multi-turn systems, risk should be assessed over complete conversational and action trajectories,
including boundary migration from informational to directive behavior. Sponsors should test semantically
equivalent prompts, changing context, contradictory inputs, long-context degradation, repeated attempts
to bypass restrictions, and transitions between modes. Under-escalation and over-escalation should be
reported separately because their harms, thresholds, and acceptable tradeoffs are not interchangeable.

2. Competency-based premarket evaluation is appropriate, but
benchmarks must be connected to deployment controls
Response to Questions 7-17. Device benchmarking followed by risk-proportionate clinical confirmation is a
sound framework. It should not, however, be treated as a one-time examination. A benchmark result is
meaningful only when the sponsor can demonstrate that the tested configuration is the configuration
deployed, that changes are controlled, and that the relevant competencies remain enforceable at runtime.

CDRH should require sponsors to define acceptance criteria at three levels:

• Outcome level: clinical correctness, safety, communication, calibration, subgroup performance, refusal,
and escalation.
• Process level: adherence to required steps, use of authorized evidence, correct tool selection,
preservation of human-review gates, and appropriate handling of missing or conflicting information.
• Control level: evidence that a safety-relevant rule can be observed, alerted upon, or enforced; that
violations generate records; and that an unsafe action can be prevented or stopped.

Public comment | Page 2
i-GENTIC AI | FDA-2026-N-7874

Public benchmarks can support development but should not be the sole gate for authorization because of
contamination, test optimization, and limited representativeness. Sponsors should establish construct
validity by linking each benchmark to the intended use, foreseeable users, clinical environment, failure
modes, and real-world input distribution. Sequestered data, independent adjudication, challenge sets, and
prospective or shadow-mode evidence should increase with consequence and autonomy.

Synthetic data may supplement rare-event, edge-case, adversarial, and coverage testing, but should
remain separately reported from real-world evidence. Sponsors should disclose the generator, generation
protocol, independence from the device under test, subgroup coverage, fidelity limitations, and the
sensitivity of results to distribution shift. Synthetic evidence should not mask weak performance for
underrepresented populations.

Independent third parties can strengthen confidence by maintaining sequestered datasets, validating
testing methods, and adjudicating results. Program design should require technical competence, conflicts
disclosure, reproducible methods, transparent qualification criteria, and portability of evidence so that
third-party participation does not create a closed market or vendor lock-in.

3. Postmarket oversight should become continuous and
event-driven for higher-risk systems
Response to Questions 18-24. Greater reliance on postmarket evidence may be appropriate only where
the residual uncertainty is bounded, the likely harm is reversible or containable, monitoring has adequate
sensitivity and timeliness, and the manufacturer can intervene before unacceptable risk materializes. It
should not substitute for sufficient premarket evidence where autonomous action could cause immediate,
serious, or irreversible harm.

Periodic re-benchmarking, clinician sampling, and degradation monitoring are valuable, but cadence alone
is insufficient. CDRH should expect both scheduled review and event-triggered reassessment. Triggering
events should include:

• a foundation-model, prompt, retrieval, tool, data-source, infrastructure, or interface change;
• a performance, calibration, refusal, escalation, subgroup, or latency threshold breach;
• a new failure mode, complaint pattern, adverse event, security vulnerability, or credible threat signal;
• a material change in intended users, workflow, clinical environment, labeling, or governing requirements;
and
• evidence that the deployed system has drifted from the authorized configuration or safety envelope.

Postmarket evidence should be traceable from the governing requirement to the implemented control,
deployed version, triggering event, resulting decision, and corrective action. This is where
machine-readable policies and automated evidence collection can reduce burden while improving
inspection readiness.

4. Supervisory agents can help, but they must be independent,
bounded, and auditable
Response to Question 20. Machine-based supervisory agents can materially improve postmarket
monitoring by evaluating actions at machine speed, detecting prohibited conditions, and creating
consistent evidence. They should not be presumed reliable merely because they use a different model.
Their authority, inputs, failure modes, and own change history require validation.

For safety-critical use, CDRH should consider the following design expectations:

Public comment | Page 3
i-GENTIC AI | FDA-2026-N-7874

• architectural separation between the acting system and the supervisory control;
• deterministic enforcement for clear, safety-critical prohibitions where feasible, with semantic methods
used for interpretation and ambiguity handling;
• independent testing against false-negative, false-positive, common-mode, prompt-injection, and
tool-output manipulation failures;
• fail-safe behavior when the supervisor is unavailable, uncertain, or detects a high-severity violation;
• tamper-evident logs and versioned evidence showing which policy and model configuration governed
each action; and
• human review and escalation paths proportional to consequence, urgency, and reversibility.

A supervisory agent should itself be treated as a safety-relevant component when the device relies on it to
maintain reasonable assurance of safety and effectiveness.

5. Foundation-model files should improve transparency but not
transfer accountability
Response to Questions 24-25. A voluntary Foundation Model Device Master File could reduce duplicative
review and improve consistency, provided it is maintained at a useful level of specificity. At minimum, it
should address model and interface versioning; supported and excluded uses; known failure modes;
safety-relevant training-data provenance at an appropriate level; healthcare-relevant evaluation results;
subgroup performance; content and refusal policies; security and tool-use constraints; audit-log
availability; change-notification commitments; rollback support; and the expected effect of updates on
downstream applications.

Voluntary files will not be sufficient where developers lack incentives to disclose or update safety-relevant
information. CDRH should also consider standardized supplier interface contracts, minimum change-notice
expectations, machine-readable version manifests, and sponsor-maintained verification tests that detect
material behavioral changes. The medical-device manufacturer must remain accountable for the safety
and effectiveness of the final configured device.

6. Agentic devices require identity, authority, and action-level
controls
Response to Question 26. Agentic AI introduces a distinct risk: the system may convert an error in
reasoning into a sequence of real-world actions before a person can intervene. Evaluation should therefore
assess the complete plan-action-observation loop, not only the final output.

CDRH should expect an agent identity and authority record, comparable to a digital passport, that binds
each deployed agent to its owner, intended use, authorized tools, permission scope, model and policy
versions, human oversight requirements, and revocation status. Each consequential tool call should be
attributable to that identity.

Additional expectations should include least-privilege permissions; allowlisted tools and data sources;
pre-action policy evaluation; explicit human approval before irreversible or high-consequence actions;
separation of planning from execution where appropriate; transaction limits; safe rollback; rate and
sequence controls; resistance to prompt injection from users, retrieved content, and tools; and an
independently testable emergency stop or kill switch.

Acceptance criteria should include not only task success, but also prohibited-action rate,
policy-compliance rate, unauthorized-tool-call rate, human-checkpoint adherence, recovery from tool
failure, safe behavior under conflicting instructions, and time to containment after a violation.

Public comment | Page 4
i-GENTIC AI | FDA-2026-N-7874

7. Recommended minimum runtime governance evidence package
To translate the paper's principles into an inspectable and least-burdensome framework, CDRH should
consider defining a minimum evidence package, scaled by risk, containing:

• a versioned map from intended use, hazards, labeling, and applicable requirements to testable controls;
• the deployed model, prompt, retrieval, tool, data, and policy configuration;
• predefined acceptance criteria and the evidence used to establish them;
• runtime decisions, exceptions, overrides, and human approvals;
• change history, drift signals, re-evaluation triggers, and corrective actions;
• subgroup, site, workflow, and environment performance where clinically relevant; and
• proof that fail-safe, rollback, revocation, and emergency-stop mechanisms were tested.

Such an evidence package would allow FDA to evaluate not merely whether a sponsor has written a
governance plan, but whether the deployed system is actually governed and whether that governance can
be demonstrated through contemporaneous evidence.

Conclusion
CDRH's discussion paper is an important and timely foundation. Its strongest feature is recognition that
GenAI-enabled medical devices require risk-proportionate evaluation across the total product life cycle.
The next step should be to make that lifecycle approach operational: benchmark the system, clinically
confirm performance, control the deployed configuration, govern consequential actions before execution,
and continuously produce evidence that the authorized safety envelope remains intact.

For generative and agentic systems, governance after an incident is too late. The regulatory framework
should encourage governance before action.

Respectfully submitted,

Dr. Zahra Timsah, PhD, MBA, MSc
Founder and Chief Executive Officer
i-GENTIC AI
Houston, Texas

Primary source reviewed: U.S. Food and Drug Administration, Center for Devices and Radiological Health, Considerations for
the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for Feedback (Aug. 2026), Docket
No. FDA-2026-N-7874.

Public comment | Page 5