Solutions by Jewel (Jewel Darlington)

Solutions by Jewel (SBJ), a global AI governance advisory firm, submits the attached comment on Considerations for the Regulation of Generative AI-Enabled Medical Devices. The comment addresses Questions 1, 5, 18, 19, 20, 21, 22, 23, 24 and 26, covering risk modifiers, conversational drift, postmarket decision conditions, machine-based supervisory agents, shared ecosystem responsibility, change control, third-party foundation-model changes and agentic devices. Every recommendation concerns governance structure; clinical performance and device safety and effectiveness fall outside the comment’s scope. Contact: governance@solutionsbyjewel.com

PUBLIC COMMENT
CONSIDERATIONS FOR THE REGULATION OF GENERATIVE AI-ENABLED MEDICAL DEVICES
Docket No. FDA-2026-N-7874
Submitted by Solutions by Jewel
Jewel Darlington, Chief AI Governance Officer & Founder
governance@solutionsbyjewel.com | 678-235-5986 | solutionsbyjewel.com
October 7, 2026

OVERVIEW
Solutions by Jewel (SBJ) is an AI governance advisory firm. SBJ designs authorization, change-control, evidence and accountability structures for
organizations that deploy consequential AI systems, including organizations in healthcare. SBJ welcomes the invitation from the Center for Devices and
Radiological Health (CDRH) to address the questions that fit each respondent's expertise.
This comment answers ten questions where governance architecture applies directly: Questions 1, 5, 18, 19, 20, 21, 22, 23, 24 and 26. SBJ leaves
Questions 2, 3, 4, 6 through 17 and 25 to respondents with clinical, statistical, human-factors and model-evaluation expertise.
Scope statement. Every recommendation in this comment concerns governance structure: authorization conditions, evidence, change control, traceability
and oversight. Clinical performance, clinical validation, and the safety and effectiveness of any specific device fall outside the scope of this comment.
Terminology. In SBJ construct names such as Continuous Authorization and Transition Authorization, authorization refers to an organization's internal
governance determination that a system may operate within defined conditions. SBJ's recommendations address the conditions a manufacturer would
prespecify and the evidence a manufacturer would maintain within FDA's existing framework for marketing authorization.
Disclosure. SBJ used Anthropic's Claude as a research and drafting tool for this comment. Anthropic is a foundation-model developer. Anthropic also
appears as a subject in SBJ's V1.0 benchmark and V2.0 Research Evidence Cohort, and the Ye et al. study cited under Questions 1 and 20 used an
Anthropic model as its monitoring agent. SBJ discloses these relationships so CDRH can weigh the foundation-model discussion under Questions 20 and
24 with full context.

SUMMARY OF RECOMMENDATIONS
1. Supervisory capability is a risk modifier. Human supervision lowers a function's risk position when the supervisor is shown to have visibility, time,
authority and competence under real operating conditions. (Q1, Q26)
2. Continued reliance on postmarket monitoring rests on prespecified decision conditions. Manufacturers define in advance the results that
trigger reassessment, the function that decides, and the response timeframe. (Q18, Q19, Q22, Q23)
3. Evidence is bound to the tested configuration. Benchmark results attach to a specific model version, prompt set, retrieval method, guardrail set,
orchestration logic and interface. A change to any of these, including one made by a third-party model developer, triggers re-examination. (Q22, Q24)
4. Operational use is part of the device's risk profile. When outputs begin driving automated workflows, or a conversation moves toward directing
action, the device's risk position changes, and that change counts as material. (Q5, Q21)
5. Machine supervisors carry their own validation and authority requirements. A supervisory agent needs independent validation, defined
intervention authority and demonstrated effectiveness at the point of correction. (Q20)
6. Agentic actions trace to their source of authority. Each action an agentic device takes links back to the authorization that bounded it, and human
checkpoints before irreversible actions are enforced at the point of execution. (Q26)

EVIDENCE BASIS AND ITS LIMITS
The constructs in this comment come from SBJ's published, frozen governance architecture.
V1.0 foundation. SBJ AI Governance Architecture V1.0 was benchmarked against six publicly documented AI deployments associated with Replit,
Anthropic, Microsoft, Salesforce, Waymo and Google DeepMind. The benchmark produced 185 applicable determinations: 160 Pass, 24 Partial, 0 Fail and
1 Insufficient Evidence, with 7 further determinations recorded as Not Applicable. V1.0 then completed Adversarial Testing Cycle 1 with zero confirmed
architectural gaps and one ambiguity, resolved through interpretation as ATC1-001 (human-operational-use drift). V1.0 includes two published prototype
environments: an agentic enterprise software-change agent and a physical autonomous warehouse transport system.
V2.0 clarifications. SBJ AI Governance Architecture V2.0 carries V1.0 forward intact and adds four adjudicated clarifications: AC-01 (Authorization
Invalidation Trigger Semantics), AC-03 (Delegation and Branch-and-Merge Provenance), AC-04 (Execution Identity and Material Version Binding) and AC05 (Evidence Sufficiency as an Authorization Condition). V2.0 addresses shared-responsibility deployment (AC-07) through the existing architecture plus
contractual, procurement, evidence-access and operating controls.
External evidence. The V2.0 Research Evidence Cohort and the V2.0 External Evidence & Source Register document the external evidence behind
V2.0. The cohort aligns with the two prototype environments:
Prototype Environment Research Evidence Cohort Domain Relevance in This Comment
Agentic AI: enterprise software-change agent Agentic / Enterprise Authorization; Security Maturity Questions 20, 24, 26
Physical AI: autonomous warehouse transport Physical AI / Industrial Execution (including Waymo, Boston Question 26, agentic control of another device
system Dynamics, Apollo Tyres with AWS)
Both environments External Assurance / Disposition (NIST NVD workflow; Questions 18, 19, 21
GovRAMP continuous assurance)

Limits of this evidence.
The benchmark and the Research Evidence Cohort contain zero clinical or medical-device cases. SBJ offers its constructs to CDRH as governance
structures that transfer from adjacent domains.
The V1.0 results describe governance-rule behavior within the tested scope.
Autonomous irreversible material financial transactions sit outside the V1.0 validated scope, as published.

Solutions by Jewel | FDA-2026-N-7874 Public Comment | Page 1 of 5
 SBJ's architecture defines the governance consequence once relevant facts are available. Detection and observability are implementation capabilities
of each manufacturer's system.
SBJ's scoring, weighting and implementation mechanics are proprietary. This comment relies on published constructs.

RESPONSES TO DISCUSSION QUESTIONS
Question 1 (Section IV): Additional dimensions for the two-axis risk framework
SBJ supports the two-axis framework as an organizing heuristic. SBJ recommends representing the four dimensions CDRH names as explicit modifiers
that move a function along the consequences axis, consistent with how CDRH already treats measurement functions and patient-facing functions. SBJ
also recommends a fifth modifier: demonstrated supervisory capability.
Reversibility and intervention window. Risk rises as the time between an output and its committed effect shrinks relative to the time a supervisor
needs to detect a problem and act. SBJ's architecture records this as a principle on committed downstream effects: once an action has produced
effects in the world, those effects persist after the system halts, so containment depends on acting before commitment. CDRH's own example of an
autonomous thrombolytic order set in a stroke workflow illustrates a short intervention window.
Downstream safeguards. A safeguard earns risk credit when it is independent of the device, holds its own authority to block or modify the action, and
has been shown to act within the intervention window. SBJ's architecture names this Independent Downstream Authority.
Time pressure. Time pressure is a property of the deployment setting that compresses the review time available to a supervisor. SBJ recommends
treating it as a multiplier on the reversibility modifier.
Traceability. SBJ recommends two levels. Output-to-source traceability links an output to its primary source materials, as CDRH describes. Decisionto-action traceability reconstructs which device version, configuration and authority produced an output and which action followed. SBJ's architecture
names this Decision-to-Action Authority Traceability, and the SBJ AI Governance Decision Record V2.0 provides a published record format for it.
Supervisory capability. The activity axis treats continuous HCP supervision as a risk moderator. SBJ recommends applying that moderation in
proportion to evidence that the supervisor can observe the relevant output, has time to act, holds authority to act and has the competence to recognize
an incorrect output under real operating conditions. Benchmark element E.4 already names automation bias as one mechanism that erodes
supervision. A 2026 controlled study of experienced developers supervising coding agents found that a monitoring agent correctly flagged 16 of 19
covert sabotage attempts, and sabotage persisted in 9 of those 16 flagged sessions (Ye et al., 2026, preprint). The study concerns software
development; SBJ cites it as transfer evidence that supervision effectiveness is an empirical quantity to measure.

Question 5 (Section IV): Risk across multi-turn conversational trajectories
SBJ recommends assessing risk at the level of the trajectory, with the highest-consequence state a realistic trajectory can reach serving as the governing
risk position for the device.
SBJ's adjudicated principle ATC1-001 supports this approach. It holds that a system's consequence and exposure include material changes in how its
outputs are consumed, acted upon, escalated, automated or embedded in downstream workflows, including cases where the system is technically
unchanged. Under that principle, a conversation that moves from general information toward a specific instruction crosses a consequence boundary, and
Transition Authorization governs whether, and under what approved conditions, the device may cross it.
Characterizing intended use. The intended use statement can define an authorized envelope: the functions, the highest level of directiveness and
the consequence classes the device is authorized to reach. SBJ's architecture calls this the Authority Envelope. Emergent conversational behavior is
then evaluated against that envelope, which gives manufacturers and reviewers a stable reference point.
Runtime boundary behavior. The device maintains a conversation-level record of its position and applies prespecified behavior at each boundary:
redirection, deferral or escalation. This aligns with the cumulative-drift testing CDRH describes under benchmark element S.2.
Deployment-site changes. When a deploying organization changes how outputs are used, for example by routing conversational outputs into an
order workflow, that change counts as material for monitoring and change-control purposes, including when the device software is identical.

Question 18 (Section VI): Accepting premarket uncertainty through postmarket monitoring
Postmarket monitoring delivers value through the decisions its observations drive. SBJ recommends that any reliance on postmarket monitoring to offset
premarket uncertainty carry five program characteristics:
1. Written authorization-supporting conditions. At market entry, the manufacturer records the evidence, configuration, intended-use envelope and
deployment conditions on which the benefit-risk determination rests.
2. Prespecified triggers. Each condition maps to observable signals and thresholds. SBJ's AC-01 states the governing rule: a material change to an
authorization-supporting condition triggers reassessment of the system's authorization state, scope or continuation.
3. Assigned response authority and timeframes. A named function holds authority to narrow, suspend or roll back the device, with a maximum time
from detection to response.
4. Demonstrated monitoring capability. Before reliance begins, the manufacturer shows the program's detection sensitivity and latency, for example
against seeded faults or historical data.
5. Preserved adverse evidence. Records of failures, interventions and rejected outputs are retained and available for review. SBJ's AC-05 adds that
evidence must be sufficient for the specific decision it supports.
The approach fits devices whose failures are detectable within the monitoring cadence and whose intervention window exceeds the combined detection
and response time. Devices whose failures can produce irreversible harm faster than that detection-and-response loop warrant the full premarket
evidence burden. SBJ's Federal Continuous Authorization Control Map applies the same structure in federal information-system settings.

Question 19 (Section VI): Postmarket approaches, cadence and triggering events
SBJ supports all three approaches CDRH describes and recommends pairing a risk-proportionate cadence with event triggers that fire immediately when
the event occurs. SBJ proposes trigger classes drawn from the evolving components CDRH lists in Section II:

Solutions by Jewel | FDA-2026-N-7874 Public Comment | Page 2 of 5
 Trigger Class Example Events
Configuration change Change to model, prompts, retrieval strategy, guardrails, orchestration logic or user interface (AC-04)
Foundation-model change Provider version change, including changes the provider deploys silently
Operational-use change New downstream integration, automation of a previously reviewed step, new user population at a site
(ATC1-001)
Environment change Shift in input population or data environment
Incident signal Material incident, complaint cluster or adverse event report
Agentic change New tool, changed tool permission or changed downstream system

SBJ also recommends three additional approaches:
Configuration attestation. Continuous verification that the deployed configuration matches the benchmarked configuration, using execution identity
and material version binding (AC-04).
Canary re-benchmarking. A fixed subset of the premarket benchmark runs automatically on every configuration event and on a frequent schedule.
Recorded determinations. Each reassessment produces a decision record identifying the trigger, the evidence relied upon, the decision made and
the accountable function.
SBJ's public interactive preview, the i-PARC™ Cockpit (SBJ Governance Cockpit), illustrates this trigger model: a model version change, environment
change, tool or permission change, or material incident each prompts reassessment of the configured authority. The i-PARC™ Cockpit is an illustrative
preview; SBJ's validation claims attach to the V1.0 benchmark. SIGNAL™, SBJ's AI-specific governance control and audit methodology, applies the same
trigger structure in periodic audit.

Question 20 (Section VI): Machine-based supervisory agents
SBJ supports supervisory agents as a monitoring component under four conditions.
1. The supervisor is a validated, version-bound component. It carries its own benchmark, acceptance criteria, configuration binding and change
control. A change to the supervisor counts as a change to the monitoring program and triggers re-evaluation.
2. The supervisor is independent. Where feasible, it comes from a model developer separate from the device's foundation-model developer, which
reduces correlated failure. This mirrors the adjudicator-independence principle CDRH already extends to LLM adjudicators in Section V.B.2.
3. The supervisor's authority is defined. The manufacturer specifies what the supervisor does on detection (record, alert, hold or block) and which
human or function receives an alert with authority to act.
4. Effectiveness is measured at the point of correction. SBJ recommends that CDRH evaluate detection rate and correction rate as two separate
measured quantities. In the Ye et al. study, the monitoring agent flagged 16 of 19 sabotage attempts, and sabotage persisted in 9 of those 16 flagged
sessions after the human supervisor received the alert. Detection and correction are distinct links in the control chain, and each needs its own
evidence.
Supervisor outputs, including misses discovered later, are preserved as adverse evidence for the monitoring program.

Question 21 (Section VI): Ecosystem roles and manufacturer accountability
SBJ recommends an explicit evidence-allocation model. SBJ's AC-07 treats shared responsibility as a deployment condition, governed through
contractual, procurement, evidence-access and operating controls. Applied to postmarket monitoring:
The manufacturer holds accountability for the device and for the design of the monitoring program.
Each ecosystem party holds defined evidence obligations: what it observes, what it records, what it reports to the manufacturer and within what
timeframe.
Healthcare institutions report operational-use changes, such as integration of device outputs into automated workflows, because those changes alter
device risk (ATC1-001).
Professional societies and standards bodies contribute reference benchmarks, adjudication criteria and reporting formats that every party can apply
consistently.
Defining each party's role as an evidence contribution keeps accountability concentrated with the manufacturer. A single test checks the arrangement: can
the manufacturer obtain sufficient authority, execution and evidence information from every party to make the determinations it is accountable for? Where
the answer is incomplete, the gap becomes a deployment condition the manufacturer resolves through contract or addresses through labeling.

Question 22 (Section VI): Scaling re-benchmarking to the nature of a change
SBJ recommends scaling evidence according to which authorization-supporting conditions a change touches. SBJ's Three-Axis System State (Lifecycle
Stage, Authorization Status and Runtime Operational State) lets a manufacturer classify each change by its effect on each axis. SBJ offers the following
tiers as an illustration of governance structure:
Tier Change Characteristics Evidence and Pathway
A. Quality system documentation Change leaves every benchmarked component identical, Record in the quality management system
confirmed by configuration attestation
B. Targeted re-benchmarking Change to one component inside the authorized envelope with Re-run the benchmark elements mapped to that
bounded expected impact component, plus safety elements S.1 to S.3, within the
quality system or a PCCP
C. Full re-benchmarking Foundation-model version change, orchestration change or new Full benchmark against premarket acceptance criteria
tool under a PCCP
D. FDA review Change to the intended-use envelope, movement up either axis, Premarket submission
new action-taking capability or new user population

Mapping each device component to the benchmark elements it can affect is a premarket design task, which reinforces CDRH's observation that the
premarket assessment should be designed with later re-benchmarking in mind. SBJ's architecture adds an operating requirement on assurance-claim
validity: when a material change affects the conditions or evidence behind an assurance claim, the claim's continuing validity is reassessed. STRATA™,

Solutions by Jewel | FDA-2026-N-7874 Public Comment | Page 3 of 5
SBJ's governance and GRC foundation, places this classification inside an organization's existing quality and risk management system.

Question 23 (Section VI): Adapting PCCPs when modifications resist prespecification
SBJ recommends that PCCPs for GenAI-enabled devices prespecify the governance envelope and the decision rules that apply to any modification. The
plan can contain:
The authorized envelope: intended-use boundaries, the highest permitted level of directiveness and the consequence classes in scope.
Component change classes with the re-benchmark subset mapped to each, as described under Question 22.
Exit rules: the changes that leave the PCCP and require a submission, written as AC-01 triggers.
Acceptance criteria tied to the premarket baseline, with safety elements held to the baseline level or better.
Rollback and Safe State: a preserved prior configuration and a defined safe operating condition for use during evaluation.
Resumption rules: after any suspension, authority to resume is issued fresh for the resulting state and conditions. SBJ's architecture names this
Derived Authorization.
A decision record for each implemented change.

Question 24 (Section VI): Changes initiated by third-party foundation-model developers
Detect. SBJ recommends three layers. First, version pinning to specific model snapshot identifiers wherever the provider offers them. Second, execution
identity binding: each device output records the model identifier, configuration and version that produced it, which makes an unannounced change visible
in the device's own records (AC-04). Third, behavioral fingerprinting: the canary benchmark subset described under Question 19 runs on a frequent
schedule and is compared statistically against the baseline.
Evaluate. Any provider-initiated model change is treated as a Tier C change under Question 22, with re-benchmarking completed before the device relies
on the new version.
Respond. When detection shows the device running on a version outside its bound configuration, the device moves to a prespecified Safe State, such as
routing to a human reviewer or pausing the affected function, for the duration of evaluation.
Mechanisms.
Contractual: advance notice of version changes, published deprecation timelines, continued access to the prior version during evaluation, incident
notification and audit log access. CDRH lists update notification commitments and audit log availability as possible Foundation Model MAF content;
SBJ recommends the same items as standard contract terms between manufacturers and providers.
Technical: version pinning, execution identity binding, behavioral fingerprinting and configuration attestation.
PCCP: a prespecified re-benchmark protocol for provider version changes, with exit rules for changes that alter the intended-use envelope.
Under AC-07, the manufacturer's ability to obtain this evidence from the provider is itself a deployment condition, to be confirmed before market entry and
maintained throughout the device lifecycle.

Question 26 (Section VII): Agentic GenAI-enabled devices
SBJ supports benchmark element A.1 and recommends four additional considerations.
1. Authority Envelope enforced at execution. The agent operates inside a bounded set of actors, actions, targets, tools, credentials, environments,
time limits and stop conditions. Agent planning, generated subgoals and retrieved instructions operate as inputs inside that envelope. Acceptance
criteria test the envelope where actions execute, through tool-level permission scoping and credential scoping. This directly supports the promptinjection resistance CDRH lists under A.1, including injection through tool outputs.
2. Delegation provenance (AC-03). Each consequential action traces back through delegation, branching and merging to its originating authorization.
The record answers five questions: who or what executed, under which authority, through which execution path, with which configuration or version,
and whether authorization held as conditions changed.
3. Checkpoint integrity. A.1 tests compliance with human-oversight checkpoints before irreversible or high-consequence actions. SBJ recommends
acceptance criteria that also test the checkpoint itself: the checkpoint sits in a component outside the agent's control, and Independent Termination
Authority sits outside the agent's own decision process. The supervisory-capability modifier from Question 1 then applies to the human at the
checkpoint.
4. Independent downstream authority. CDRH notes that an agentic system's functions could meet the device definition, for example when its action
sequences result in control of another medical device. In that configuration, the downstream device holds its own authority checks and its own defined
Safe State. SYNTHESIZE™, SBJ's physical AI governance methodology (PERCEIVE, DECIDE, ACT, OBSERVE, LEARN), governs this class of
physical action. The supporting evidence comes from the Physical AI / Industrial Execution domain of the V2.0 Research Evidence Cohort and is
transfer evidence for medical settings.
To reflect elevated risk in acceptance criteria, SBJ recommends scaling them with three measurable properties: the number of consequential actions per
task executed between checkpoints, the reversibility of those actions, and the breadth of tool permissions. Where agentic operation shortens human
review, risk credit for supervision follows demonstrated supervisory capability. For postmarket oversight, agentic devices add the trigger classes listed
under Question 19: new tools, changed permissions and changed downstream systems.

CLOSING
SBJ appreciates CDRH's early, open engagement on generative AI-enabled devices. The competency-based approach gives the field a strong premarket
structure. The governance structures in this comment are offered to help that premarket evidence stay valid as devices, their foundation models and their
deployment settings change. SBJ is available to discuss any part of this comment.

RESPONDENT INFORMATION

Solutions by Jewel | FDA-2026-N-7874 Public Comment | Page 4 of 5
RESPONDENT INFORMATION
Organization Solutions by Jewel
Point of Contact Jewel Darlington, Chief AI Governance Officer & Founder
Email governance@solutionsbyjewel.com
Phone 678-235-5986
Website solutionsbyjewel.com

SCOPE AND INTELLECTUAL-PROPERTY NOTE
Where an SBJ construct is cited, it is cited as a documented architecture concept publicly retrievable at the locations provided in this comment. SBJ's
proprietary implementation internals, scoring architecture, weighting and internal assessment criteria remain protected and sit outside this comment.
Application of these constructs to GenAI-enabled medical devices is Solutions by Jewel's professional technical position, offered for CDRH's independent
evaluation.
Validation limitation. Autonomous, irreversible material financial transactions sit outside the scope of V1.0 adversarial validation, and the validation
results speak to the execution contexts that were tested.
SBJ AI Governance Architecture, STRATA™, SIGNAL™, SYNTHESIZE™, Decision Record and the i-PARC™ Cockpit are original Solutions by Jewel
work. External people, organizations and systems named in SBJ's evidence records are independent evidence sources that SBJ evaluated on its own
initiative.
This comment contains only information SBJ intends for the public record, with zero proprietary, confidential or personally identifiable information.

REFERENCES
1. U.S. Food and Drug Administration, CDRH. Considerations for the Regulation of Generative AI-Enabled Medical Devices: Discussion Paper and Request for
Feedback. August 18, 2026. https://www.fda.gov/media/194242/download
2. U.S. Food and Drug Administration, CDRH. Executive Summary for the Digital Health Advisory Committee Meeting: Total Product Lifecycle Considerations for
Generative AI-Enabled Devices. November 2024. https://www.fda.gov/media/182871/download
3. U.S. Food and Drug Administration, CDRH. Marketing Submission Recommendations for a Predetermined Change Control Plan for Artificial Intelligence-Enabled
Device Software Functions. Guidance, as cited in reference 1, footnote 22.
4. Ye, Zou, Yu and Shi. Coding with "Enemy": Can Human Developers Detect AI Agent Sabotage? arXiv:2606.05647, version 1, June 4, 2026. Preprint.
https://arxiv.org/abs/2606.05647
5. Solutions by Jewel. SBJ AI Governance Architecture V1.0. 2026. https://solutionsbyjewel.com/sbj-ai-governance-architecture-v1
6. Solutions by Jewel. SBJ AI Governance Architecture V2.0. 2026. https://solutionsbyjewel.com/ai-governance/architecture/
7. Solutions by Jewel. V2.0 Research Evidence Cohort. 2026. https://solutionsbyjewel.com/ai-governance/architecture/assets/sbj-ai-governance-architecture-v2research-evidence-cohort.pdf
8. Solutions by Jewel. V2.0 External Evidence & Source Register. 2026. https://solutionsbyjewel.com/ai-governance/decision-record/sbj-ai-governance-architecturev2-external-evidence-source-register.pdf
9. Solutions by Jewel. SBJ AI Governance Decision Record V2.0 and technical package (JSON Schema 2.0.0). https://solutionsbyjewel.com/ai-governance/decisionrecord/ | https://github.com/solutionsbyjewel/sbj-ai-governance-decision-record
10. Solutions by Jewel. Federal Continuous Authorization Control Map. https://solutionsbyjewel.com/federal-continuous-authorization-control-map
11. Solutions by Jewel. i-PARC™ Cockpit (SBJ Governance Cockpit), interactive governance preview. https://solutionsbyjewel.com/

Solutions by Jewel | FDA-2026-N-7874 Public Comment | Page 5 of 5