Krishna Chaitanya Kotabhattara

Re: Docket No. FDA-2026-N-7874.

Comment on Question 20, machine-based supervisory agents

Submitted by Krishna Chaitanya Kotabhattara, MS, Data Science.

I am an engineer at ConcertAI. I write in my personal capacity and these views are my own, not my employer’s.

I built a machine-based supervisory agent of the kind Question 20 describes, tried to measure its reliability the cheap way, and found that the cheap way cannot work. I am writing because the reason it cannot work applies to anyone who builds one, and because it is easy to miss.

The cheap way, and why it fails

A supervisory agent watches a device’s output and flags what looks wrong. To trust the agent you need to know how often it misses something it should have flagged. The obvious way to find out is to plant known errors in its queue and count how many come back caught. This is what seeded probes, synthetic test cases and injected faults all do.

The number a probe returns cannot stand in for the number you need, for three reasons I measured.

First, the error runs one way only. Cases the reviewer marked safe are the cases that go into the record, so anything missed is filed as safe. The recorded risk contracts to at most one minus the miss rate of its true value. A monitoring scheme built this way never reads worse than reality. It reads safer, always, and by an amount nobody can see from inside the system.

Second, the correction does not apply. Every published method for correcting a rate measured under noisy labels requires a condition on how the contaminated and clean score distributions relate. I simulated four reviewer behaviours, from attentive to rubber-stamping, and the condition failed in all four, in the direction that voids the correction rather than the direction that makes it conservative.

Third, the quantity is not identified. Two deployments whose reviewers miss defects at different rates can return the same planted-error score, because planted defects and the defects a model produces on its own are not equally hard to notice. No fixed conversion factor between the two exists, and I could find no study that measures both on the same clinical material.

What I suggest CDRH ask for instead

I do not suggest that supervisory agents be excluded. I suggest that four things be expected of any claim about one.

1. Adjudicated labels for any stated miss rate. A second independent reader on a sample of real cases settles what a probe cannot. I priced it rather than asserting it is affordable, and the price depends sharply on how attentive the reviewer already is. The number of adjudicated cases needed scales as the negative log of the significance level divided by the square of the gap being detected. Under the contamination model in my paper that is 185 cases for a reviewer who misses 85 percent of defects, about 2,100 for one who misses a quarter, and 147,938 for one who misses 3 percent. The better the reviewer, the more expensive it is to prove it, which is a fact a monitoring requirement should face rather than assume away.

2. An interval on every estimate, which is already CDRH’s stated expectation elsewhere, since inferences from point estimates ignore the uncertainty in them.

3. The right words for agreement. When a supervisory agent is compared against another model, an unadjudicated panel or a synthetic label, the comparator is not a reference standard, and CDRH’s own diagnostic-test guidance says sensitivity and specificity are not appropriate terms in that case. Positive and negative percent agreement are. The same guidance warns against dropping equivocal results, which for a supervisory agent means that timeouts, refusals and unparseable outputs belong in the denominator.

4. Fail-closed behaviour, stated and tested. An outage, a timeout or an unmet precondition must be reported as an inconclusive result and never as a clean one. This is a property a reviewer can test directly, and it is cheap to check. The library I work on refuses to apply the correction above until the score condition is confirmed on adjudicated data, and returns an explicit inconclusive verdict when it cannot confirm it.

What I am not claiming

Every reviewer number in this work is simulated. No clinician was measured. The direction of the result follows from the algebra and needs no measurement, and only the constants come from simulation. The empirical grounding is other people’s published clinical data, read one study at a time, and I deliberately quote no average across those studies, because the effects disagree in sign and averaging them would manufacture the very conversion factor I argue does not exist.

Availability

The full argument, the simulation and the code are in a manuscript I am preparing for publication. Every number in it is computed by a named function. The attached letter is this comment in full.

Respectfully submitted,

Krishna Chaitanya Kotabhattara, MS, Data Science
kkrishnachaitanya@concertai.com

Re: Docket No. FDA-2026-N-7874; Considerations for the Regulation of Generative AI-Enabled
Medical Devices: Discussion Paper and Request for Feedback

To the Center for Devices and Radiological Health:

Comment on Question 20, machine-based supervisory agents
Submitted by Krishna Chaitanya Kotabhattara, MS, Data Science.
I am an engineer at ConcertAI and I write in my personal capacity. These views are my own and not my
employer’s. I name the affiliation only because it is where the work described happened.
CDRH asks whether postmarket monitoring can be facilitated by machine-based supervisory agents, and
what considerations apply, including the evaluation and reliability of the supervisory agent itself. I built
such an agent, tried to measure its reliability the cheap way, and found that the cheap way cannot work. I
am writing because the reason it cannot work applies to anyone who builds one, and because it is easy to
miss.

The cheap way, and why it fails
A supervisory agent watches a device’s output and flags what looks wrong. To trust the agent you need to
know how often it misses something it should have flagged. The obvious way to find out is to plant known
errors in its queue and count how many come back caught. This is what seeded probes, synthetic test cases
and injected faults all do.
The number a probe returns cannot stand in for the number you need, for three reasons I measured.
First, the error runs one way only. Cases the reviewer marked safe are the cases that go into the record, so
anything missed is filed as safe. The recorded risk contracts to at most one minus the miss rate of its true
value. A monitoring scheme built this way never reads worse than reality. It reads safer, always, and by an
amount nobody can see from inside the system.
Second, the correction does not apply. Every published method for correcting a rate measured under noisy
labels requires a condition on how the contaminated and clean score distributions relate. I simulated four
reviewer behaviours, from attentive to rubber-stamping, and the condition failed in all four, in the direction
that voids the correction rather than the direction that makes it conservative.
Third, the quantity is not identified. Two deployments whose reviewers miss defects at different rates can
return the same planted-error score, because planted defects and the defects a model produces on its own
are not equally hard to notice. No fixed conversion factor between the two exists, and I could find no study
that measures both on the same clinical material.

What I suggest CDRH ask for instead
I do not suggest that supervisory agents be excluded. I suggest that four things be expected of any claim
about one.
1. Adjudicated labels for any stated miss rate. A second independent reader on a sample of real cases settles
what a probe cannot. I priced it rather than asserting it is affordable, and the price depends sharply on how
attentive the reviewer already is. The number of adjudicated cases needed scales as the negative log of the
significance level divided by the square of the gap being detected. Under the contamination model in my
paper that is 185 cases for a reviewer who misses 85 percent of defects, about 2,100 for one who misses a
quarter, and 147,938 for one who misses 3 percent. The better the reviewer, the more expensive it is to prove
it, which is a fact a monitoring requirement should face rather than assume away.
2. An interval on every estimate, which is already CDRH’s stated expectation elsewhere, since inferences
from point estimates ignore the uncertainty in them.
3. The right words for agreement. When a supervisory agent is compared against another model, an unadjudicated panel or a synthetic label, the comparator is not a reference standard, and CDRH’s own diagnostictest guidance says sensitivity and specificity are not appropriate terms in that case. Positive and negative
percent agreement are. The same guidance warns against dropping equivocal results, which for a supervisory
agent means that timeouts, refusals and unparseable outputs belong in the denominator.
4. Fail-closed behaviour, stated and tested. An outage, a timeout or an unmet precondition must be reported
as an inconclusive result and never as a clean one. This is a property a reviewer can test directly, and it
is cheap to check. The library I work on refuses to apply the correction above until the score condition is
confirmed on adjudicated data, and returns an explicit inconclusive verdict when it cannot confirm it.

What I am not claiming
Every reviewer number in this work is simulated. No clinician was measured. The direction of the result
follows from the algebra and needs no measurement, and only the constants come from simulation. The empirical grounding is other people’s published clinical data, read one study at a time, and I deliberately quote
no average across those studies, because the effects disagree in sign and averaging them would manufacture
the very conversion factor I argue does not exist.

Availability
The full argument, the simulation and the code are in a manuscript I am preparing for publication. Every
number in it is computed by a named function, and the code, the score sets and the record of every run are
deposited at doi:10.5061/dryad.g79cnp66x. That deposit is held private until publication, and I will supply
it to CDRH on request before then.
Respectfully submitted,
Krishna Chaitanya Kotabhattara, MS, Data Science
kkrishnachaitanya@concertai.com