Novamigo (Matthew Grecsek)

Novamigo submits the attached comment on Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical Devices. We ask FDA to let patient-facing AI tools go as far as their errors can be counted: a lane, with postmarket evidence as its primary assurance, for tools that record every expectation, opinion and direction before the outcome, are graded by what happened next, and report those counts; and a time-limited evaluation period open to any patient-facing tool that can report them. The attached letter answers Questions 1, 2, 3, 5, 6, 11, 14, 15, 18, 19, 21, 24 and 25.

Comment on Docket No. FDA-2026-N-7874

Novamigo
Matthew Grecsek · founder · matt@novadenovo.com

Dockets Management Staff, Food and Drug Administration
Re: Docket No. FDA-2026-N-7874, Considerations for the Regulation of Generative AI-Enabled Medical
Devices: Discussion Paper and Request for Feedback

Summary

We ask FDA to let patient-facing AI tools go as far as their errors can be counted. A tool that records every
expectation, opinion and direction before the outcome, is graded by what happened next, and reports
those counts to FDA should have a lane that relies on postmarket evidence as its primary assurance, and
should leave that lane when its counts show harm. We also propose a time-limited evaluation period, open
to any patient-facing tool, general-purpose chatbots included, that can report those counts.

A medical record is never read on its own: a clinician reads it with judgment about which values matter, for
which condition, against which plan. A chatbot placed over the record takes that judgment out and puts
nothing standard in its place: no standard for which model answers, how the question was asked, or which
context it reads the record in. That is the risk. A computed record works the other way: it gives the clinician,
and the patient, more context than the record holds, so the judgment applied to it is better informed.

Patients are already asking AI what their own medical records mean and acting on the answers: 15 percent
of US adults use AI chatbots to help decide whether to go to a doctor (Pew Research Center, 2026). The
largest AI companies now connect patients' records to general-purpose models, and each that states an
intended use disclaims diagnosis and treatment, so the largest exposure to patient-facing clinical AI sits
outside oversight while a tool built for the purpose carries the burden. The incentive runs against the safest
designs. Patients rely on FDA for safety, and postmarket surveillance is a mechanism FDA has long used and
proved; it cannot work on advice no one records.

We ask FDA to state that lane in the final framework, for patient-facing tools that meet six conditions:

1. every output traceable to the patient's own record;
2. every finding about the case computed from the record, and any inference from medical knowledge
anchored to those findings, labeled as inference and logged;
3. every option the record supports shown with its basis and the strength of its evidence, low grades
included and labeled, and the decision left with the patient and their clinician;

Novamigo · October 19, 2026 1 / 8
 4. any opinion on which option the analysis favors, to the patient, the clinician or both, stated as an
opinion with its basis, evidence and uncertainty open; any direction given only on a decision path
placed for that class of situation, such as seeking care now;
5. every expectation, opinion and direction recorded before the outcome and graded by it, counting errors
in both directions;
6. those counts reported to FDA and published, de-identified, and a tool whose counts show harm leaves
the lane.

The lane would move what is already happening, at national scale, into the one place where its errors are
counted, published and corrected, in view of FDA, clinicians and patients.

Who we are and what Novamigo does

Novamigo is a service for patients, with any condition, offered directly to them. For every patient it turns the
record they already have into a running picture of their health: not a file of documents, but how the case is
changing against its plan, with no new tests and nothing new for the care team to install. For a patient
living with disease, it shows whether their care is still on its plan. For the most serious cases, it shows the
turns that matter while the decision is still open. In cancer, by the most conservative published anchors,
tens of thousands of deaths a year in the United States are excess deaths, winnable with treatment that
already exists, chosen differently or in time (Hanna et al., BMJ 2020; Newman-Toker et al., BMJ Quality and
Safety 2023). We are not promising cures. Our work is to make those treatments visible while there is still
time to use them.

Every plan is written for a population, and no one can say in advance whether this patient, with their own
biology, their own disease and their own response to treatment, will stay inside the population their plan
assumes. Novamigo assembles the record from every institution that holds a piece of it and follows each
new lab, scan and note as it arrives. At every step it recontextualizes the whole case and asks whether, at
this moment, the case still fits the plan. It writes down what the plan expects before each result arrives, so a
departure is caught while it is still an option; most days, it confirms the plan is working. Every finding about
the case comes from a deterministic core and traces to the record. On that anchor, a language model
reasons from medical knowledge to explain the record, show the options on the table and weigh them, and
what it infers is labeled and logged.

I built Novamigo from my wife's stage 4 cancer case, and in it I saw the benefits and the risks of AI in
patient care firsthand. Novamigo has run alongside that care: 466 source documents from 6 institutions
over 18 months. It found at least six turns, each dated and shown while the decision was still open. Where
one was acted on, the trajectory reversed. Oncology is where Novamigo is proven; the method is general,
because any condition with a plan has an expectation the next result can confirm or contradict.

The premise: the current framework is not reaching what patients already do

Today, patients ask AI as a matter of course what their own medical records and symptoms mean, and act
on the answers, with no oversight of those answers and no count of when they are wrong. That is at odds
with what device regulation, as FDA currently expresses it, is meant to do, and it needs to change. Here is
why.

Novamigo · October 19, 2026 2 / 8
Usage is measured. In a June 2026 survey, 34 percent of US adults said they had used AI chatbots for at
least one health purpose: 25 percent to figure out what is causing symptoms, 20 percent to understand lab
results, and 15 percent to help decide whether to go to a doctor (Pew Research Center, 2026). In a survey
fielded in early 2026, 13 percent of US adults had uploaded personal medical information such as test
results or doctors' notes to an AI chatbot (KFF, 2026).

The record connection is now a product. OpenAI's ChatGPT Health connects records through b.well;
OpenAI reports that "over 230 million people globally ask health and wellness related questions on
ChatGPT every week," and states that Health "is not intended for diagnosis or treatment." Amazon's Health
AI, "your personalized health agent," retrieves records through health information exchange, "reviews your
conditions, medications, and allergies before answering," and offers to "turn symptoms into actionable
outcomes"; Amazon states it "is not intended for diagnosis or treatment without the support of a care
provider." Anthropic, Microsoft, Perplexity and Google offer the same through third-party record networks,
and Epic's Emmie answers patients from their own chart inside MyChart at health systems that have
adopted it (each as of October 2026).

Connected is not anchored. These tools help, but they are navigators: they answer about what already
happened, from what they are shown. A record holds values at points in time. A clinician reading it brings
the judgment that turns those values into a case: which ones matter, for which condition, against which
plan. A model reading the record brings neither the plan nor that judgment, so it sees less than a visit
averaging twenty minutes does (Shaw et al., American Journal of Managed Care 2014), the way a
photograph sees a moving car. A few highly capable patients build that anchor themselves, with a team
around them. Most cannot. An answer from an unanchored picture can move a patient toward the paper's
own concerns, "delaying needed care or pursuing inappropriate self-treatment," and toward error in both
directions: care that was not needed, and care that was.

The disclaimer is the boundary, and FDA is already asking whether a disclaimer constrains anything.
The discussion paper considers whether a patient-facing function "may not become any less directive
because it includes a 'talk to your doctor' or an 'I am not a medical professional' statement." Yet under
intended use, a statement of that kind places these tools outside device oversight altogether. Either these
are device functions, and the largest exposure to patient-facing clinical AI is unregulated, or they are not,
and FDA should say where the line falls so that a tool built for the purpose can rely on it too.

What the status quo costs. Failure of care delivery, failure of care coordination and overtreatment or lowvalue care waste an estimated $205 billion to $345 billion a year in the US (Shrank et al., JAMA 2019). That
is the part anyone can count. It stands in for what no one counts: the suffering of patients whose care went
wrong in ways a computed account of their case, read with clinical knowledge, could have shown while the
decision was still open.

The other side of the ledger. The paper names it: "the democratization of clinical information," "patient
empowerment and engagement," and "avoiding medical parentalism that may underestimate patient
capability." We agree. These benefits are reachable when the inference stands on a computed, auditable
account of the case. That is the lane we ask for, and the answers below describe it.

Risk

Q1. Does the two-axis framework capture the dimensions most relevant to risk?

Novamigo · October 19, 2026 3 / 8
Add three dimensions, and let consequence and context decide.

• Traceability to the patient's own record. The paper treats a function as higher risk when the user
"cannot independently evaluate the basis for the output." A patient who can open the page each finding
comes from can evaluate its basis, and so can the clinician the patient brings it to.
• Whether inference is anchored. A general-purpose chatbot reading the record inherits the record's
gap, not the clinician's judgment. A patient without clinical training usually cannot fill the gap, because
they cannot know which points are relevant. The context chosen decides the answer: for a patient with
diabetes and cancer, a rising glucose reads as poor diabetes control against one plan and as the effect
of the steroids given with chemotherapy against the other, and reading it against the wrong one
misleads. An inference from medical knowledge carries far less risk when it stands on a computed,
checkable account of this patient's whole case, every condition and every plan in it, than when it reasons
from whatever it is shown.

A computed record, not an assembled one. Assembling the record does not anchor anything: a model
reading an assembled record still sees values at points in time, without the judgment a clinician brings to
them. A computed record adds what the values hold only in motion: how each is moving, how fast,
against the others, and against the plan. It does more than stand in for that judgment. It gives the
clinician, and the patient, more context than any one reader of the record has, and that is what separates
it from a chatbot reading the record.

• Who holds the next action. Where a licensed prescriber stands between an output and any action, the
consequence of an error is moderated; a patient cannot write a prescription, order a lab test or a scan, or
start chemotherapy. Where the patient acts alone (going to the emergency room or not, an over-thecounter remedy, stopping a medicine), there is no downstream safeguard. That is the decision 15
percent of US adults now take to a chatbot.
• Consequence and context decide what is allowable, and accountability decides how far a tool may
go. For a low-consequence, reversible question (an upset stomach and what people commonly take for
it), an answer that names the usual remedies is technically a recommendation and should not be a
device function. Where consequence is high, context decides the risk: an answer drawn from an
incomplete picture of the case is more likely to be wrong, whether a clinician or an AI gives it, and there
a wrong answer costs the most. Set consequence against context and the risk concentrates in one cell,
high consequence with poor context, which is where general-purpose chatbots now weigh options for
patients from whatever they are shown. Accountability is the second test: whether anyone can count
how often the answer was right.

Q2. How should directiveness along a continuum modify risk?

For high-consequence patient-facing functions, let context and accountability move the line along the
continuum, not wording. The risk of a directive output has two parts: resting on poor context, it is likely to
be wrong, and without a count, no one can measure whether it was. The paper's second step ("in situations
similar to this, clinicians often increase the lisinopril dosage") is information, and the clearest case of it is the
record compared with the plan the patient's own physician wrote: two facts in the patient's own record,
measured against each other. A tool that commits its expectations in advance and has its calls graded
should also be able to state which option its analysis favors and why, to the patient as well as the clinician,
as an opinion with its basis, evidence and uncertainty open. Biology is uncertain, so that opinion never says

Novamigo · October 19, 2026 4 / 8
a course is safe or will work; it states what the evidence shows about benefit, harm and uncertainty for each
option, and the decision stays with the patient and their clinician. Directing a patient to a specific treatment
action stays out of the lane, except on a placed decision path. A tool that cannot show the context its
opinion rests on, or how often its opinions were right, should stay at the second step.

Regulate decision paths, not sentences. Many patient situations resolve into a few paths, each with a cost
if it is wrong: watch and re-measure within a stated window; see your own clinician within a stated time;
seek care now. FDA could place the paths for a class of situation on the consequence axis, hold a tool's
outputs to them, and audit after market how each turned out. Where a path is placed, a tool may direct a
patient to it, and a manufacturer knows in advance what its tool may say. Every direction is graded by what
followed, so a tool that sends too many patients to emergency care, or too few, shows in the counts. An
illustration for one class, with the thresholds and warning signs left for FDA or a clinical society to set:

Class: home blood pressure What the tool may say Moves to the next path Cost if this path is wrong
well above the patient's usual when
range, no symptoms

1. Watch and re-measure Re-measure at set times over a readings stay above the a short delay
stated window, and report the stated threshold at the
readings end of the window

2. See your own clinician soon Contact your clinician within a a warning sign appears a visit
stated number of days, bringing
the readings and the record

3. Seek care now Go to emergency care now an unneeded emergency
visit, against a missed
stroke or heart attack

Each time, the tool records the path it chose, the facts behind the choice, and what happened next.

An example from my own family. Abroad, my father's blood pressure, normally well controlled, rose to a
concerning level and stayed there for several days. No local doctor knew his record, and his own physician
could not be reached. A careful conversation with a general-purpose AI advised against raising his dose (it
would take weeks to act, and could leave his pressure too low once he was home) and proposed watching
for 24 to 48 hours and reporting new readings. The readings came down, and his physician later confirmed
each step. It worked because the model was capable and I knew how to give it context. Most patients have
neither, and no one counted whether that answer was right. A tool held to the same paths, working from
the patient's actual record, does this for a patient who cannot, and records how it turned out.

Q3. When does delivering clinical information to patients present different risks, and what safeguards
help without underestimating patients?

The risk to patients today is asymmetric information. Care runs on it, most of all in complex disease, and the
gap is wider far from an academic center. A physician knows more medicine than the patient and can still
reach a wrong judgment for want of the whole context. A patient who holds that context brings it into the
room, and the decision made with the physician rests on more. That is patient empowerment in the sense
the paper uses: context in the patient's hands, so that their decisions, made with their clinicians, rest on the
whole picture. Where the record supports options beyond those the care team has raised, the tool shows
them too, and in that sense it can act as a second opinion; its purpose is to bring those options into the
conversation with the care team, not to stand against it. The safeguards are the lane's: every finding

Novamigo · October 19, 2026 5 / 8
traceable, every option shown with the strength of its evidence, the next step framed as a question for the
clinician, and physician review available on request. A safeguard should not withhold. Where the evidence is
even, as when an option helps about as often as it does not, the choice belongs to the patient, so the tool
shows the option with its evidence grade, a low grade labeled as low, rather than deciding for the patient
by leaving it out. This matters most for the patients the paper describes as facing "structural barriers to
healthcare access": a patient treated anywhere can see whether their case is still on its plan.

Q5. How should risk be assessed across conversations that may drift from information to direction?

The risk in a chatbot today is that the context under the conversation is nebulous at best: whatever the
model was shown, with no record of what it said. The remedy is a deterministic, record-based core as the
context, a language model working on top of it, and every turn logged for audit; most of the drift this
question describes comes from lacking the anchor, the audit, or both. In Novamigo the conversation is held
to what the record and the analysis show: it states no opinion the analysis did not form and gives no
direction a placed path does not allow, and every turn is logged, so a departure from either is caught. The
same principle separates two kinds of autonomy in agentic tools: gathering records, re-running the
comparison and telling the patient unasked is autonomy in attention; every action that touches care passes
through a person. Autonomy in action is a different risk class and should be assessed as one. Directiveness
is judged at every turn, where a harmful direction is given, and across the whole conversation, where drift
accumulates; a tool that keeps no record of the conversation can do neither.

Q6. How should under-escalation and over-escalation be characterized and weighed?

Count both, against what actually happened next. Every alert system, a clinician's or a machine's, can err in
two directions: raising a concern that proves unnecessary, or staying silent when it should have spoken.
Today neither direction is routinely counted for patient-facing AI. The two are not commensurable, as the
paper notes: an unnecessary visit costs time, money and worry, and a missed escalation can cost a life,
which is a reason to lean toward the visit where the patient holds the next action and to measure how often
it proved unnecessary. In Novamigo every call (raise it or stay silent) is written down before the next event
and graded by it: a catch, a false alarm, a missed alarm or a correct silence. Its accuracy is measured, not
claimed. Two rules keep those counts honest. A silence over an interval with no data is not a correct silence;
it is reported as uncovered. And a tool that stays silent when the record shows a departure from the plan
has under-escalated, whatever the departure's cause, a departure in the care itself included. The tool states
what the record shows; judging it is for others.

Evidence

Q11. How might clinical confirmation be selected and proportioned, and are there other methods?

Build the confirmation into the product. Each expectation is committed before the next event and graded
by it, so the evaluation cannot be fitted after the fact: a prespecified prediction protocol with objective
grading by the record's own future, running continuously.

• The treating physician is the validator. The physician does nothing extra: the next event in the record,
the decision they make and the result that follows, grades the expectation written before it. Each event
validates the one before. Medicine runs these experiments every day and the results go uncollected.
Novamigo collects them because that is where better treatment is found: among those results are better
outcomes and fewer side effects, and a decision that did better than its plan expected becomes

Novamigo · October 19, 2026 6 / 8
 evidence for the next patient in the same situation. The clinicians' calls are counted as evidence about
decisions and what followed, never as a judgment of the clinician.
• Every count reports how complete it is. An outcome the count cannot see is reported as unseen, never
counted as nothing having happened.

Q14 and Q15. Comparators, and performance against the absence of the device.

For a tool whose calls are logged and graded, with the decision left to the patient and their clinician,
evaluate the patient and clinician working with it, the paper's "human-AI team," against an independent
specialist panel at the standard of care. The real comparator is what the patient has today: a record split
across portals, nobody comparing it with the plan, and a chatbot whose answers nobody checks. Measure
process rather than survival: whether a departure from the plan was visible, to whom, and how long before
the next decision.

Postmarket evidence as the primary assurance

Q18, with Q19 and Q21. When is greater reliance on postmarket monitoring appropriate, and how
should it work?

When a tool's outputs are traceable, its opinions and directions are logged, and its errors are counted in
both directions: the lane's conditions. Postmarket surveillance is how FDA has long kept watch over devices
in use, and it needs a record to audit. Where a tool's advice is not recorded, its errors cannot be counted,
and harm no one counts is harm no one corrects. A record that a tool stayed within its bounds is not a
count of whether it was right; the lane asks for both. The tool does not grade itself: each call is graded by
what the record shows next, kept where the model cannot revise it, and a count of zero is reported with its
coverage, never as evidence of safety. A tool in the lane would commit to:

• FDA access to audit what the system produced and what it told patients, de-identified;
• the two counts from Q6, reported to FDA and published;
• periodic independent clinician review of samples;
• re-benchmarking against logged cases after any change.

Novamigo commits to all four, with every published finding de-identified and governed by the patient's
consent. Publication lets societies, institutions and patients check the tool, and lets patients and clinicians
choose the tools they rely on by how often each was right, while accountability stays with the manufacturer.
A tool whose counts show harm leaves the lane and returns to review. Because one tool's error repeats
across every patient it serves, the counts are reviewed at fixed intervals from the first patient, not after harm
accumulates. The lane fits where errors can be seen: where the next event in the record grades each call,
and where any irreversible act, a prescription or a procedure, passes through a clinician first. An output
whose error would never surface in the record belongs on the premarket path.

An evaluation period. Within its existing authority, or as a pilot that informs Congress, FDA could open a
time-limited evaluation period for patient-facing tools, general-purpose chatbots included, with one
condition of entry: every opinion and every direction logged, graded against what happened next, and the
counts reported. A tool in it would have wider latitude than the premarket path allows, and a tool whose
counts show harm would leave it for review. At the end, FDA would hold measured performance, in both
directions, for the tools patients already use.

Novamigo · October 19, 2026 7 / 8
Foundation models

Q24 and Q25. Third-party model changes, and Foundation Model MAFs.

Keep the model at the interface. Where the findings about a case come from the record and a deterministic
core, a change to the underlying model can change wording and inference, not the findings. Novamigo
logs what the model infers, so the inference layer is re-benchmarked against logged cases before any new
model is used. Voluntary Foundation Model MAFs would be useful, and we would reference one.

Closing

Patients will keep asking AI about their own records. The paper's own goals, fewer patients delaying needed
care or pursuing inappropriate self-treatment and more of the benefit of information in their hands, are
best served by moving that activity to where its errors are counted. Today the greatest source of risk is
general-purpose AI making high-consequence recommendations to patients from their records without the
context a clinician brings: no clinical judgment, no plan to read the record against, no standard for how it is
done, and no count of when it is wrong. Bringing it inside a counted lane is the change that matters most.
We ask FDA to state the lane described above, with postmarket evidence as its primary assurance, and to
open the evaluation period to every patient-facing tool that can report its counts. Better picture. Better
judgment. Better outcomes.

Matthew Grecsek
Founder, Novamigo
matt@novadenovo.com

Novamigo · October 19, 2026 8 / 8