AI Engineer World's Fair 2026
Inside 847 Production Clinical AI Notes — Sebastian Fox, Composo
Read the talk
Inside 847 Production Clinical AI Notes
The title points to a corpus of 847 production notes, while the recording also discusses a separate real-world study and a newly generated multi-scribe dataset without establishing that they share the same denominator. Sebastian Fox uses those sources to show why plausible clinical notes conceal dangerous omissions—and why evaluation must learn continuously from expert judgment.
From a talk by Sebastian Fox
At a glance
Ideas worth remembering
The title’s 847 notes, the cited real-world study, and the generated notes used for judge comparisons are not established as one dataset. Keep their denominators separate.
A fluent output can be dangerous because of what it omits; checking whether every written sentence is plausible does not establish completeness.
Ambient-scribe failures occur in transcription and generation, but additions, changes, and omissions become difficult chiefly when an evaluator must decide which difference matters.
A second language model is not automatically a safety net. Static-rubric judges can compare words accurately while missing intent, temporal decisions, and case-specific significance.
Store moving, explainable judgment as retrievable cases, expert corrections, and references rather than relying only on a fixed rubric or repeatedly retrained weights.
Begin by collecting expert comments on real outputs, then continuously discover failure modes, capture reasoning, and calibrate each evaluation against relevant prior cases.
The most dangerous note may look completely fine
A clinical note describes an ordinary new headache: likely tension-type, take paracetamol, return if it does not settle. Nothing on the page looks obviously false. But the consultation also included jaw pain while chewing in a patient over fifty. Fox, a medical doctor by background who now runs Composo, identifies that combination as a red flag for giant cell arteritis—a same-day emergency because untreated disease can threaten sight within days. The symptom that changes the case never reached the note.
This is a harder failure than visible nonsense. A spectacularly bad note might invent chest pain, angina, diabetes medication, and a nonexistent hospital address for a young patient with tonsillitis. Such additions attract attention. A fluent note that quietly removes the decisive fact can remain in the record because every sentence it contains is plausible.
The title identifies 847 production clinical AI notes, but the recording does not define that corpus or say that every result comes from it. Fox separately cites what he calls the largest real-world study, then later describes notes generated across three production scribes and reused for a judge comparison. Those datasets should not be assumed to be identical.
For the real-world study, Fox reports that about one in twenty notes carried a potentially seriously harmful error, nearly one in five had an important omission, and more than one in ten had a hallucination. He also says ambient scribes operate in about a third of US practices and that most systems lack adverse-event reporting. The recording does not provide the study design needed to assess how broadly those rates generalize.
That reporting gap changes the meaning of apparent safety. An unreported incident count does not show that deployment is harmless when errors can simply remain embedded in records. Healthcare makes the stakes visceral, but the same structure appears anywhere a fluent output can omit the fact, clause, or decision that controls what happens next.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Correct words can still produce the wrong meaning
Notes generated across three production ambient scribes produced a revealing distribution of failures. Fox plots each error by clinical importance on one axis and whether a strong automated check catches it on the other. A few errors are caught; the troubling cluster is both consequential and missed. Better prose quality has moved the problem away from easy-to-spot gibberish and toward subtle distortions of intent.
One patient cannot say whether her headache began suddenly or gradually; she only says that it “just happened.” The generated note converts that uncertainty into “abrupt sudden onset.” The observable change is one phrase, but its consequence is large: sudden onset is a red flag for bleeding in the brain. The model did not copy a stated fact. It resolved ambiguity, inserted the stronger interpretation, and made that interpretation drive the workup.
Another consultation discusses tests before the patient asks to try antibiotics instead. The doctor and patient agree to defer testing, treat, and observe. The note nevertheless says the tests were arranged that day. The proposed plan really did appear in the conversation, so simple transcript matching finds supporting words. What the note loses is the conversation’s temporal structure: proposal, negotiation, rejection, final decision.
Failures can enter at two layers. Transcription can confuse soundalike drug names, reverse hyperthyroidism into hypothyroidism, or drop the “no” from “no evidence of cancer.” Even with a perfect transcript, generation can still fail in three parallel ways:
- Addition: Insert something that was never said, ranging from a mild inference to a dangerous fabrication.
- Change: Rephrase supplied information in a way that may preserve or reverse its meaning.
- Omission: Remove material ranging from small talk to a clinically decisive allergy or red flag.
Detecting a difference is therefore only the mechanical half of evaluation. The evaluator must decide which difference matters in this case. Fox calls that “taste,” meaning domain judgment rather than aesthetics. It is tacit because experts cannot fully enumerate it, contextual because the same detail changes importance across cases, and moving because models, guidelines, institutions, and expert opinions change. A model may know the medical facts while still lacking a case-specific sense of priority.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
A checker can become a second silent failure
The natural response is to separate generation from verification. Let one model write the note, then give a second model more time, tokens, context, and a narrow instruction to find what is wrong. This asymmetry works especially well in mathematics and code, where a compiler, unit test, or executable answer can cheaply reject many bad outputs.
A serious clinical judge can receive the transcript, generated note, surrounding context, a detailed faithfulness rubric, and worked pass/fail examples. Teams can optimize the rubric and add deterministic NLP checks that count medical concepts appearing in one document but not the other. Expert review remains the offline gold standard, but the automated judge is supposed to extend coverage to every production note.
Fox built that version and ran the same multi-scribe notes through it. The judge flagged a handful and approved most; one in five of its clean passes still contained a serious error, often an omission. This is a result from the presented judge comparison, not necessarily from all 847 notes or from the separate real-world study. It exposes the central failure: a capable model can compare two documents without knowing which missing or changed fact deserves priority.
Where does the safety net break? The flow below makes the two silent decisions visible. The generator can preserve words while misrepresenting intent, and the checker can approve that note because its static standard does not rank the case’s decisive difference highly enough.
Contains symptoms, proposals, uncertainty, negotiation, and final decisions.
Matching words is insufficient when the final decision or important omission is the information that controls safety.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The same omission can be noise or the answer
Verification is cheap when correctness has an external test. Clinical completeness has no equivalent of a compiler. A judge can spot differences between transcript and note, but it must also construct the standard that tells it which differences matter. A predefined rubric contains only the judgment its authors managed to express in advance; the hardest part is the judgment they could not fully write down.
Two patients have blood in their urine, and both notes omit where they recently travelled. One went to France; the other went to Lake Malawi. The omissions have the same textual shape. In the first case, the holiday detail is irrelevant. In the second, freshwater exposure in sub-Saharan Africa points toward schistosomiasis and changes the management plan. A concept-counting check sees two missing travel locations. Clinical judgment sees noise in one and a diagnostic clue in the other.
That leaves three places to store the standard. A prompt or rubric is inspectable and easy to change, but it freezes only what people can specify. Model weights can absorb examples through fine-tuning or continual learning, but Fox considers them poorly suited to a moving, explainable standard because updates require retraining and the resulting score cannot point cleanly to the judgment that changed it. Retrieved examples preserve past decisions, expert corrections, and references outside the weights, then bring the relevant subset into the judge’s context for each output.
The third option treats expert judgment as cases rather than a complete specification. A new correction can affect the next evaluation call immediately, and reviewers can inspect which examples or references moved the score. This is a retrieval and context-assembly problem: find prior cases with the same meaningful shape, not merely the same words.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Discover, capture, and calibrate the standard continuously
Fox organizes the alternative as a repeating three-step loop: discover failures in real outputs, capture how experts judge them, and calibrate each new output against the cases and references that bear on it. The standard remains editable because it lives in data and context rather than only in a fixed prompt or retrained model.
Discover from production. Cluster observed failures and name the recurring modes to build a failure-mode ontology. Synthetic cases remain useful, but they cover failures someone already imagined. Real outputs reveal unexpected behavior. The ontology is not a checklist the judge blindly applies; it organizes what experts should review and how prior cases should be indexed for retrieval.
Capture reasoning, not just scores. Put real outputs before clinicians for focused review and collect comments, corrections, and explanations. A useful seed can come from a few concentrated hours rather than a months-long labeling project. The resulting record shows how experts actually draw distinctions, including distinctions they would struggle to turn into universal rules.
Calibrate per output. Keep generic instructions such as “be faithful” or “do not omit important information,” then assemble the case-specific standard at evaluation time. Retrieve similar judged outputs, applicable expert corrections, reference documents, and guidelines. Each new judgment and correction becomes material for later evaluations; newly discovered failure modes feed back into the same system.
What relationship does this loop make visible? Evaluation produces both a decision and new memory. The next judge call is therefore informed by the failures and corrections accumulated before it, while discovery keeps the retrieval scheme from being limited to an ontology guessed at launch.
Supply failures that synthetic tests may not anticipate.
Real failures determine what experts review; expert reasoning becomes retrievable context; each evaluation produces more cases for the next round.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
The missing red flag becomes visible when the judge knows what matters
Return to the opening headache. The evaluator need not have memorized that patient. It can retrieve earlier cases with the same shape—a red flag summarized as routine—plus expert corrections about new headaches in patients over fifty and applicable criteria or guidelines. Once that context sets the priority, the omitted jaw pain stands out against an otherwise plausible note.
Fox compares three judges on the same generated-note dataset: an off-the-shelf model with a rubric, a more engineered judge with rubric optimization and deterministic checks, and a judge supplied with discovered failure modes and per-output expert context. The third performs better on this specific dataset. The recording supplies no result showing that this comparison covered the full 847-note corpus; its useful lesson is narrower. The model and notes stay the same while the evidence and judgment standard placed in context change.
The pattern extends beyond medicine. A contract reviewer can omit the clause that changes the deal, while a support agent can promise a refund the company does not offer. Both may be watched by a generic judge that detects obvious inconsistency but lacks the organization’s sense of consequence.
The smallest useful starting point is also the most human: ask experts to leave free-form comments on real outputs. Those comments are raw material for discovering failure modes, preserving corrections, creating retrieval indexes, and eventually automating more of the review. Evaluation cannot be built once and frozen when the standard itself changes with cases, policy, and experience. It is an operating loop, not a possession.
Suggest correction
This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.
Resources
From the talk
Composo’s site links its failure-mode taxonomy, clinical omission work, benchmarks, and papers on improving language-model judges—the most direct supplied resource for investigating the methods behind the presentation.
Related talks
- The Future of Evals: From LLM as a Judge to Agent as a Judge — Aparna Dhinakaran, Arize AI
Extends the discussion from fixed-rubric judges to evaluators that inspect traces and discover unexpected production failures.
- The maturity phases of running evals
Explains how expert rationales can become automated scoring functions and how production clustering helps teams discover new failure modes.
- AI-powered entomology: Lessons from millions of AI code reviews
Shows the same evaluation problem in code review: models can identify many possible issues, but undocumented team judgment determines which feedback developers actually want.
Read the complete timestamped transcript
- 0:01
[music]
- 0:12
>> This is a clinical note an AI wrote from
- 0:15
a real consultation.
- 0:17
Take a few seconds and read it.
- 0:19
It reads like a routine headache.
- 0:21
A new headache, likely tension type,
- 0:23
take some paracetamol, come back if it
- 0:25
doesn't settle.
- 0:27
Looks completely fine, doesn't it?
- 0:30
Here's what's missing.
- 0:32
In the room, she also mentioned her jaw
- 0:35
aches when she chews. A new headache,
- 0:37
over 50 with jaw pain on chewing,
- 0:40
that's giant cell arthritis.
- 0:42
And untreated, it can take her sight
- 0:43
within days. It's a same day start
- 0:46
steroids now emergency.
- 0:48
And that one line, it never made it into
- 0:49
the note.
- 0:50
On the page, it's a paracetamol
- 0:51
headache.
- 0:52
And nothing in the note is technically
- 0:54
wrong.
- 0:55
It's the dangerous part is what isn't
- 0:58
there.
- 0:59
And so that's what I'm going to talk
- 1:00
about today.
- 1:01
The dangerous failures are often the
- 1:03
ones that actually look completely fine.
- 1:07
Firstly,
- 1:08
who am I? I'm Seb, medical doctor by
- 1:10
background, and now I'm Composure, where
- 1:13
we build AI evaluation systems for
- 1:16
high-stakes domains.
- 1:19
So, that one was a subtle kind of error,
- 1:22
but sometimes it's not subtle at all.
- 1:25
A man in his 20s sees his GP for a sore
- 1:27
throat, tonsillitis.
- 1:29
The AI writes that up. It gets him chest
- 1:32
pain, suspected angina, diabetes
- 1:34
medications he's never taken, and an
- 1:36
address for a hospital that doesn't
- 1:37
exist.
- 1:38
And I I really like the LLM for this
- 1:41
one. I think it's it's a good attempt at
- 1:42
hospital name. Um
- 1:45
and weeks later, he's invited to
- 1:46
diabetic eye screening for diabetes he
- 1:48
doesn't have.
- 1:50
That's genuinely a real case that
- 1:51
happened recently.
- 1:54
Obviously, these kind of crazy ones
- 1:55
someone notices, but it's those quiet
- 1:58
ones that sit in the record uncalled
- 2:01
that are the most challenging
- 2:03
and can actually do a lot more damage.
- 2:06
And they're not rare at all. In the
- 2:09
largest real-world study of these notes,
- 2:11
about 1 in 20 carried an error that was
- 2:13
serious enough that it could cause
- 2:15
significant harm to the patient.
- 2:17
1 in 20. That's not theoretical in
- 2:19
testing, that's in production on real
- 2:21
patients.
- 2:23
And that's only the serious the ones. If
- 2:25
you widen that lens to all errors,
- 2:27
nearly 1 in 5 had an important omission
- 2:30
and more than 1 in 10 had a
- 2:31
hallucination.
- 2:33
And AI is being deployed at scale across
- 2:36
healthcare fast. Ambient Scribes are one
- 2:39
of the leading cases, already in about a
- 2:41
third of US practices and climbing.
- 2:43
Physician AI use doubled last year and
- 2:46
none of this is tracked.
- 2:48
So, for most of these systems, there's
- 2:49
no adverse event reporting at all. The
- 2:52
errors never show up as incidents, they
- 2:54
just sit in the record.
- 2:55
So,
- 2:57
errors this common that are going unseen
- 2:59
is is quite hard for me to believe that
- 3:01
it's not already affecting patients.
- 3:03
It's not that we checked and it's fine,
- 3:05
it's that we're flying blind.
- 3:09
And
- 3:10
this isn't just a healthcare problem,
- 3:12
it's every high-stakes use of AI.
- 3:14
Healthcare shows it more viscerally
- 3:16
because here being confidently wrong
- 3:18
can be life and death.
- 3:20
But, everything I show you can map
- 3:21
straight back onto other domains as
- 3:23
well.
- 3:25
So, here's what I want to do. I'm going
- 3:27
to show you what exactly is going wrong,
- 3:29
why it's going wrong,
- 3:31
why the systems we built to catch it
- 3:32
don't work, and
- 3:35
a suggestion at how maybe we can start
- 3:37
to fix that.
- 3:39
So, first, what's going wrong and why?
- 3:41
So,
- 3:42
LLMs are getting good, obviously. They
- 3:44
don't make stupid mistakes anymore most
- 3:46
of the time. So, it's not about dumb
- 3:47
errors.
- 3:48
Everything here came out of three of the
- 3:50
best production Ambient scribes on the
- 3:52
market.
- 3:54
Ones that we all know.
- 3:55
We generated a load of notes across them
- 3:57
last week. And this is exactly what's
- 4:00
going on right now. This is every
- 4:02
failure we found. Each dot is an error
- 4:04
colored by type.
- 4:06
Left to right, how much it matters.
- 4:07
Bottom to top, whether a strong
- 4:09
automated check catches it.
- 4:11
And that split is the point.
- 4:14
A handful up top get caught. But almost
- 4:16
everything sits below the line. The ones
- 4:19
I care about most are these on the
- 4:21
bottom right. The high stakes and missed
- 4:24
ones.
- 4:26
Let me show you what a couple of those
- 4:27
looks like.
- 4:29
So, a woman comes in with a headache.
- 4:33
Doctor asks,
- 4:34
"Did it come on suddenly or build up
- 4:36
gradually?"
- 4:37
She says she doesn't know. It just
- 4:38
happened. The note records that as
- 4:40
abrupt sudden onset.
- 4:42
And sudden onset is a red flag. You can
- 4:44
see why it just happened could maybe be
- 4:48
interpreted and inferred as abrupt
- 4:49
onset.
- 4:50
But, that's a feature that points to a
- 4:52
bleed on the brain. She never said it.
- 4:54
The model decided it. And now that one
- 4:56
word drives the whole workup.
- 5:00
Here's another.
- 5:01
Doctor suggests running some tests.
- 5:03
Patient says, "Can we just try try
- 5:05
antibiotics instead?"
- 5:07
They agree, hold off on the tests,
- 5:09
treat, and see how it goes.
- 5:11
Note records the opposite.
- 5:13
Arrange tests today.
- 5:15
It kept the plan that they talked out
- 5:17
of, not the one they chose. Every line
- 5:19
in the note reads fine because it's not
- 5:22
really a hallucination at all. It's not
- 5:24
wrong. It was there in the original.
- 5:26
But, it's just not what they ended up
- 5:27
deciding.
- 5:30
So, why are these happening?
- 5:33
There's, you know, in ambient scribes,
- 5:35
there's first transcription and then
- 5:37
generation.
- 5:38
A lot of it
- 5:39
does happen on the transcription layer.
- 5:42
It can be words misheard for their
- 5:44
sound-alikes. So, Humalog heard Humulin.
- 5:48
Two insulins on completely different
- 5:50
timelines, so swapping them could crash
- 5:52
a blood sugar.
- 5:53
Hyperthyroidism becomes hypothyroidism,
- 5:56
the opposite condition.
- 5:57
Or a drop to no
- 6:00
on uh no evidence of cancer that becomes
- 6:03
evidence of cancer.
- 6:04
So, these these are really hard
- 6:05
problems, and they are common.
- 6:07
Not the ones I'm going to focus on,
- 6:09
because
- 6:10
most of what goes wrong is actually even
- 6:12
with a perfect transcript.
- 6:14
It's the model reading the words
- 6:16
correctly and still doing one of three
- 6:18
things.
- 6:19
Either it adds something that was never
- 6:21
said, it changes something that was, or
- 6:24
it omits something that should be there.
- 6:28
Now, the blatant version of each of
- 6:30
these is is really easy to catch. The
- 6:32
hard part in all three
- 6:34
is
- 6:36
the same. It's telling whether that
- 6:38
thing that was added or changed or
- 6:39
dropped actually matters.
- 6:41
It's detecting that slight over
- 6:43
inference versus the dangerous
- 6:45
fabrication. The harmless rephrase
- 6:47
versus the meaningful edit.
- 6:49
A dropped line of small talk versus a
- 6:51
dropped allergy.
- 6:53
So, the ones that matter slip through
- 6:55
along with all of the ones that don't.
- 6:59
That call which different matters
- 7:02
is taste, effectively.
- 7:05
Not aesthetic taste, but essentially
- 7:08
judgment. It's It's whether in this
- 7:10
context a missed allergy might kill
- 7:12
someone or is not important.
- 7:15
And I think there's there's three
- 7:16
properties that really matter about
- 7:17
this.
- 7:18
It's tacit, so your domain experts have
- 7:21
it, but they can't fully write it down.
- 7:23
It's contextual, so the same detail is
- 7:25
critical in one note, noise in the next.
- 7:28
And it's moving. The model changes,
- 7:30
guidelines change, two good doctors
- 7:32
disagree, different hospitals have
- 7:34
different definitions. So, there's no
- 7:36
fixed target to write down.
- 7:39
And so, the model knows the facts,
- 7:42
ultimately. They're extraordinarily
- 7:43
capable, but what they lack is a sense
- 7:46
of what matters
- 7:47
here, for this specific example. And
- 7:49
that's why even brilliant models make
- 7:51
these mistakes.
- 7:53
So,
- 7:54
one natural move, you're never going to
- 7:56
make that generator perfect. Generator
- 7:58
is cheap. Generation is cheap, so stop
- 8:00
fixing it at the source.
- 8:02
Let it write, put a checker after it,
- 8:04
pass only what clears the bar.
- 8:06
And that checker should be the easier
- 8:08
job. The generator has to get everything
- 8:10
right and it pay attention to lots of
- 8:13
varying instructions.
- 8:15
Whereas the checker only has to find the
- 8:16
one thing that's wrong and just focus on
- 8:18
that task. You can also give it more
- 8:19
time, more tokens, the exact failure
- 8:21
modes to hunt for.
- 8:23
Evaluation should be easier than
- 8:24
generation.
- 8:25
It's the asymmetry of verification,
- 8:28
verifies law. That's why AI is raced
- 8:30
ahead anyway, you can cheaply check the
- 8:32
answer, maths and code.
- 8:35
And
- 8:36
doing this is exactly what the best
- 8:38
teams do. They put a lot of energy into
- 8:39
evaluation. It starts with the gold
- 8:42
standard, which is expert humans
- 8:44
reviewing notes, which obviously works
- 8:46
offline, but you can't put a human on
- 8:48
every note in production.
- 8:50
So, they automate it.
- 8:52
They build a
- 8:54
serious system and
- 8:56
some of the best versions of this that
- 8:57
I've seen are
- 9:00
you take the transcript and the note
- 9:02
and context, put in front of the judge,
- 9:05
a detailed rubric for faithfulness with
- 9:07
worked, pass and fail examples.
- 9:10
The rubric maybe auto-optimized with GPA
- 9:12
or something like that. Maybe you have
- 9:13
some deterministic NLP to sort of count
- 9:16
up medical concepts that are differing
- 9:18
between the two.
- 9:20
That's a powerful system.
- 9:22
And yet, I pulled all of those errors
- 9:24
earlier
- 9:25
out of Ambient Scribes in an afternoon.
- 9:29
So, if the evaluation is this good, how
- 9:31
are these errors still getting through?
- 9:34
So,
- 9:35
I built this system and ran those same
- 9:38
notes through it.
- 9:39
And it scored most of them fine.
- 9:42
It flagged a handful of them
- 9:44
and signed off the rest.
- 9:46
>> [clears throat]
- 9:47
>> But one in five of those clean passes
- 9:49
still had some sort of serious error
- 9:51
buried in it.
- 9:52
And often that was an omission.
- 9:55
The things that should have been there
- 9:57
and actually quietly weren't.
- 9:59
And that's the best version of a judge
- 10:02
I've seen in a lot of teams and it waved
- 10:04
them through.
- 10:06
Why did it do that? It's not stupid.
- 10:09
It's a frontier model, serious
- 10:10
engineering behind it, more than clever
- 10:12
enough to read the whole encounter and
- 10:14
catch every obvious error. And it's not
- 10:16
blind, either.
- 10:17
And and that's part of the trap. If you
- 10:19
take a note that says start amoxicillin,
- 10:21
when the real decision was actually to
- 10:23
wait and see,
- 10:25
it's faithful to the words, amoxicillin
- 10:27
did come up, but it's a lie about the
- 10:29
intent.
- 10:30
A good judge might catch that, might.
- 10:34
But whether it flags that versus the
- 10:36
other
- 10:37
doesn't have other things that it could
- 10:38
comment on
- 10:39
depends on it knowing what decision
- 10:41
matters most.
- 10:42
And so
- 10:44
it's not blind, it just can't tell what
- 10:46
counts, essentially.
- 10:48
So the note passes confidently and you
- 10:50
put a judge like that in front of your
- 10:52
system, you've not added a safety net,
- 10:54
you've added a second silent failure
- 10:56
that just nods along with the first.
- 10:59
And here's the root of it. So in math or
- 11:02
code, the verifier comes with free, a
- 11:05
unit test, a compiler.
- 11:08
But for is this note safe and complete,
- 11:10
there's no unit test.
- 11:12
You have to build the verifier yourself
- 11:14
and verification is only easier than
- 11:16
generation for the easy bit, i.e. spot
- 11:20
the difference between transcription
- 11:21
note.
- 11:22
But that's not the hard bit. The hard
- 11:23
bit is knowing of all those differences
- 11:25
you've seen, which matter. And that's
- 11:26
harder than writing that plausibly good
- 11:28
general note in the first place.
- 11:31
Because that standard of good was never
- 11:33
written down anywhere that the judge can
- 11:34
read it.
- 11:36
A rubric that you pre-specify is only
- 11:38
the taste you could write down.
- 11:40
The taste that matters is the part that
- 11:42
you couldn't.
- 11:43
And so here's here's a bit more detail
- 11:45
on what what matters looks like.
- 11:48
Two patients, both with blood in their
- 11:50
urine, both notes dropped the same kind
- 11:52
of line where they'd been on holiday.
- 11:55
One had been to France, the other to
- 11:56
Lake Malawi. Same English emission, same
- 11:58
shape, same mistake.
- 12:01
Well, not really, because blood in the
- 12:04
urine obviously warranted away
- 12:06
and you're going to have to investigate
- 12:07
it, but the France trip is irrelevant.
- 12:10
The Lake Malawi trip is the diagnosis.
- 12:12
Fresh water in sub-Saharan Africa means
- 12:14
schistosomiasis until proven otherwise
- 12:17
and it completely changes what the
- 12:18
management plan is. So that same dropped
- 12:20
line in one note is pure noise, in the
- 12:23
other it's the answer.
- 12:24
And which one it is, you simply just
- 12:26
can't write all of that down in advance.
- 12:29
So
- 12:30
if you can't write it down, you can't
- 12:31
write taste down, how do you get that
- 12:33
into your evaluator and your whole
- 12:36
application system?
- 12:38
Well,
- 12:39
we've answered a version of this before.
- 12:41
RLHF exists because you can't write the
- 12:43
reward function for good. You learn it
- 12:45
from examples by showing it.
- 12:47
The only question is where you keep what
- 12:49
you've learned.
- 12:50
And there's three places.
- 12:52
You can either specify it up front, you
- 12:54
can
- 12:55
stuff the prompt, write the perfect
- 12:56
rubric. We just watched that fail
- 12:59
essentially.
- 13:01
You can bake into the weights,
- 13:03
fine-tuning or continual learning, but
- 13:06
for a standard that's still moving and a
- 13:07
score that has to be explainable,
- 13:09
the weights, I think, are the wrong
- 13:11
place to keep that.
- 13:13
They go stale, they can't tell you why
- 13:16
and
- 13:17
you can't change them without a retrain.
- 13:19
So there's the third option, which I'll
- 13:21
show you, which is you essentially just
- 13:22
keep the taste as the examples
- 13:24
themselves. Past judgments, expert
- 13:26
corrections, references, and for each
- 13:28
output, you retrieve the ones that bear
- 13:30
on it into the judges context,
- 13:33
add one and it's live on the next call.
- 13:35
You can point at exactly what moved the
- 13:36
score.
- 13:38
For this problem, it's both better and
- 13:40
also cheaper to do.
- 13:44
So, that's the way to do that is one
- 13:46
repeating loop, three steps.
- 13:48
Discover the failure modes from real
- 13:50
outputs, capture how your experts judge
- 13:52
them, calibrate every output against
- 13:55
that, and when the standard moves, the
- 13:56
loop moves with it.
- 13:59
So, in more detail, discover. You don't
- 14:01
write that rubric in a vacuum.
- 14:03
You have to put the system in production
- 14:04
and look at the real outputs.
- 14:06
Cluster what goes wrong and the failure
- 14:07
modes surface on their own. You name
- 14:09
them.
- 14:10
This is your failure mode ontology.
- 14:12
Discover from your data, not guess on a
- 14:14
whiteboard.
- 14:15
And you can't shortcut it. The ways that
- 14:17
a real system goes wrong are effectively
- 14:19
unbounded and synthetic test cases only
- 14:22
cover the failures you already imagined.
- 14:25
The ones that hurt you are often the
- 14:26
ones that you didn't. And you'll only
- 14:28
find those in real outputs.
- 14:30
So, this ontology is your map, what to
- 14:32
capture judgment on, what to retrieve
- 14:34
against, including the failures that you
- 14:36
never thought to check for.
- 14:39
After that, it's capture and then
- 14:41
calibrate. So, those discovered modes,
- 14:44
they're not a checklist that the judge
- 14:45
runs, but they organize everything. What
- 14:47
you What you ask your experts about, how
- 14:49
you index the cases that you'll
- 14:50
retrieve,
- 14:52
and
- 14:53
capturing is a simple part. You put real
- 14:55
outputs in front of your experts.
- 14:57
Clinicians spend a focused few hours
- 14:59
leaving comments. A session doesn't have
- 15:01
to be a month-long labeling project to
- 15:03
start with. And you collect their
- 15:04
judgment.
- 15:06
Not just a score, but the reasoning and
- 15:07
corrections. And over time, you build up
- 15:09
that record of how your experts actually
- 15:11
judge.
- 15:12
You then calibrate.
- 15:14
That's
- 15:15
the the the generic part of this you can
- 15:16
write down once easily. For example, be
- 15:18
faithful or don't drop anything
- 15:20
important.
- 15:22
But what you can't write down is what
- 15:23
counts as a serious miss for this
- 15:25
specific note. That's contextual. And it
- 15:27
shifts from note to note.
- 15:29
So,
- 15:31
what we recommend is you assemble that
- 15:33
on the fly. For each output, your
- 15:35
judging agent pulls in everything that
- 15:37
bears on this one case. It's memory of
- 15:39
the most similar outputs
- 15:41
that it's judged before and how they
- 15:42
scored, the expert corrections that
- 15:44
apply, the reference documents and
- 15:45
guidelines.
- 15:47
Just context engineering per output.
- 15:50
And
- 15:51
crucially not just one pre-specified
- 15:53
rubric in a vacuum, and not a model that
- 15:55
you have to retrain every week,
- 15:57
but a full sort of case-specific
- 15:59
standard assembled for this output.
- 16:03
And it's a loop as well.
- 16:05
Every output you judge, every
- 16:06
correction, sharpens the next.
- 16:09
And when a brand new failure mode
- 16:10
appears, Discovery surface it, and it
- 16:12
flows straight back in.
- 16:15
And so, to make that a little bit more
- 16:15
concrete,
- 16:17
that headache that I opened with, the
- 16:19
one that was really a possible blindness
- 16:20
emergency,
- 16:22
here's the kinds of things that you
- 16:23
would want to pull in for that note.
- 16:25
The nearest cases that your experts have
- 16:27
judged,
- 16:28
not this exact patient, but the same
- 16:30
shape, maybe a red flag filed as
- 16:32
routine.
- 16:33
Uh the corrections that apply, like a
- 16:35
new headache over 50, um
- 16:38
suggests something that you need to
- 16:40
check red flags on, and some criteria
- 16:43
and guidelines, and you pull all of that
- 16:45
in.
- 16:46
It hasn't memorized this case. It's a
- 16:48
capable model.
- 16:50
And handed the right context to reason
- 16:51
from, held against that, the dropped red
- 16:53
flag stands out. It was never actually
- 16:55
hard to catch. It just didn't know what
- 16:57
mattered.
- 16:59
And so, if you take that same data set
- 17:00
generated notes from the start and pass
- 17:02
it through these three
- 17:04
judging systems,
- 17:06
the first, a strong off-the-shelf
- 17:09
judge with a rubric frontier model,
- 17:11
um
- 17:13
it's better than a coin flip, but it
- 17:14
misses most of what matters.
- 17:17
The second, that sort of serious system
- 17:18
that we talked about before,
- 17:20
rubric, deeper,
- 17:22
maybe some detona stick checks, better
- 17:24
again,
- 17:26
but still missing quite a lot of what
- 17:27
counts.
- 17:28
The third, the judge running this loop,
- 17:30
discovered failure modes, calibrated for
- 17:32
output against what experts judged,
- 17:35
is
- 17:36
performing a lot better on this specific
- 17:38
data set.
- 17:39
Same notes.
- 17:41
The only thing that changes is what the
- 17:42
judge was shown.
- 17:43
And the difference here, it's not more
- 17:44
compute or a better prompt, it's that
- 17:46
the first two fight taste and lose. They
- 17:48
guess the criteria, they freeze one
- 17:50
standard, and they go stale.
- 17:53
This repeating evolving loop does the
- 17:55
opposite. It discovers the modes, fits
- 17:57
the standard to each mode, and keeps
- 17:59
learning.
- 18:01
So, you might not write chemical notes,
- 18:04
but if you ship anything where being
- 18:06
confidently wrong has a cost, the
- 18:08
contract review that misses the clauses
- 18:11
that change the deal, the support agent
- 18:12
that promises a refund you don't offer,
- 18:16
the same thing is true for all of those.
- 18:17
It's watched, if at all, by a judge with
- 18:21
no taste for what matters in your
- 18:22
domain. So,
- 18:25
three things. Discover your failure
- 18:26
modes from real outputs, don't guess
- 18:28
them. Capture your experts' judgment on
- 18:31
them, the standard that they can't write
- 18:33
down.
- 18:34
Calibrate every output against the cases
- 18:36
that they've already judged, not a
- 18:38
static rubric, not a retrained model.
- 18:40
Then keep that loop running.
- 18:42
And if you take one thing away,
- 18:44
easiest place to start is your experts
- 18:47
leaving free-form comments on real
- 18:49
outputs.
- 18:50
That's the real That's the raw material
- 18:52
for everything else.
- 18:54
Your judge can verify anything that you
- 18:57
write down in advance, but the standard
- 18:59
of good never could be. And so, stop
- 19:01
trying to write it all down in advance,
- 19:03
and just start capturing it case by
- 19:05
case, and evolving it.
- 19:07
That's why evaluation can't be a thing
- 19:09
you build once and freeze.
- 19:11
The standard it checks against doesn't
- 19:14
exist on paper.
- 19:15
It has to be discovered from real
- 19:16
outputs captured from the people who
- 19:18
hold it and kept alive as it moves.
- 19:22
Evaluation isn't something you have,
- 19:23
it's something that you do continuously
- 19:25
over time.
- 19:27
Thank you.
- 19:29
>> [applause]