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.

0:120:30
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

0:12 · section reference included

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.

3:334:03
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

3:33 · section reference included

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.

How it fits togetherHow a plausible note and its checker can fail together

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.

7:598:29
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

7:59 · section reference included

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.

11:0011:31
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

11:00 · section reference included

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.

How it fits togetherThe discover–capture–calibrate evaluation loop

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.

13:4414:14
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

13:44 · section reference included

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.

16:1516:45
Suggest correction

This note stays in this page until you copy or download it. Nothing is submitted; reloading clears the draft.

16:15 · section reference included

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.

Read the complete timestamped transcript
  1. 0:01

    [music]

  2. 0:12

    >> This is a clinical note an AI wrote from

  3. 0:15

    a real consultation.

  4. 0:17

    Take a few seconds and read it.

  5. 0:19

    It reads like a routine headache.

  6. 0:21

    A new headache, likely tension type,

  7. 0:23

    take some paracetamol, come back if it

  8. 0:25

    doesn't settle.

  9. 0:27

    Looks completely fine, doesn't it?

  10. 0:30

    Here's what's missing.

  11. 0:32

    In the room, she also mentioned her jaw

  12. 0:35

    aches when she chews. A new headache,

  13. 0:37

    over 50 with jaw pain on chewing,

  14. 0:40

    that's giant cell arthritis.

  15. 0:42

    And untreated, it can take her sight

  16. 0:43

    within days. It's a same day start

  17. 0:46

    steroids now emergency.

  18. 0:48

    And that one line, it never made it into

  19. 0:49

    the note.

  20. 0:50

    On the page, it's a paracetamol

  21. 0:51

    headache.

  22. 0:52

    And nothing in the note is technically

  23. 0:54

    wrong.

  24. 0:55

    It's the dangerous part is what isn't

  25. 0:58

    there.

  26. 0:59

    And so that's what I'm going to talk

  27. 1:00

    about today.

  28. 1:01

    The dangerous failures are often the

  29. 1:03

    ones that actually look completely fine.

  30. 1:07

    Firstly,

  31. 1:08

    who am I? I'm Seb, medical doctor by

  32. 1:10

    background, and now I'm Composure, where

  33. 1:13

    we build AI evaluation systems for

  34. 1:16

    high-stakes domains.

  35. 1:19

    So, that one was a subtle kind of error,

  36. 1:22

    but sometimes it's not subtle at all.

  37. 1:25

    A man in his 20s sees his GP for a sore

  38. 1:27

    throat, tonsillitis.

  39. 1:29

    The AI writes that up. It gets him chest

  40. 1:32

    pain, suspected angina, diabetes

  41. 1:34

    medications he's never taken, and an

  42. 1:36

    address for a hospital that doesn't

  43. 1:37

    exist.

  44. 1:38

    And I I really like the LLM for this

  45. 1:41

    one. I think it's it's a good attempt at

  46. 1:42

    hospital name. Um

  47. 1:45

    and weeks later, he's invited to

  48. 1:46

    diabetic eye screening for diabetes he

  49. 1:48

    doesn't have.

  50. 1:50

    That's genuinely a real case that

  51. 1:51

    happened recently.

  52. 1:54

    Obviously, these kind of crazy ones

  53. 1:55

    someone notices, but it's those quiet

  54. 1:58

    ones that sit in the record uncalled

  55. 2:01

    that are the most challenging

  56. 2:03

    and can actually do a lot more damage.

  57. 2:06

    And they're not rare at all. In the

  58. 2:09

    largest real-world study of these notes,

  59. 2:11

    about 1 in 20 carried an error that was

  60. 2:13

    serious enough that it could cause

  61. 2:15

    significant harm to the patient.

  62. 2:17

    1 in 20. That's not theoretical in

  63. 2:19

    testing, that's in production on real

  64. 2:21

    patients.

  65. 2:23

    And that's only the serious the ones. If

  66. 2:25

    you widen that lens to all errors,

  67. 2:27

    nearly 1 in 5 had an important omission

  68. 2:30

    and more than 1 in 10 had a

  69. 2:31

    hallucination.

  70. 2:33

    And AI is being deployed at scale across

  71. 2:36

    healthcare fast. Ambient Scribes are one

  72. 2:39

    of the leading cases, already in about a

  73. 2:41

    third of US practices and climbing.

  74. 2:43

    Physician AI use doubled last year and

  75. 2:46

    none of this is tracked.

  76. 2:48

    So, for most of these systems, there's

  77. 2:49

    no adverse event reporting at all. The

  78. 2:52

    errors never show up as incidents, they

  79. 2:54

    just sit in the record.

  80. 2:55

    So,

  81. 2:57

    errors this common that are going unseen

  82. 2:59

    is is quite hard for me to believe that

  83. 3:01

    it's not already affecting patients.

  84. 3:03

    It's not that we checked and it's fine,

  85. 3:05

    it's that we're flying blind.

  86. 3:09

    And

  87. 3:10

    this isn't just a healthcare problem,

  88. 3:12

    it's every high-stakes use of AI.

  89. 3:14

    Healthcare shows it more viscerally

  90. 3:16

    because here being confidently wrong

  91. 3:18

    can be life and death.

  92. 3:20

    But, everything I show you can map

  93. 3:21

    straight back onto other domains as

  94. 3:23

    well.

  95. 3:25

    So, here's what I want to do. I'm going

  96. 3:27

    to show you what exactly is going wrong,

  97. 3:29

    why it's going wrong,

  98. 3:31

    why the systems we built to catch it

  99. 3:32

    don't work, and

  100. 3:35

    a suggestion at how maybe we can start

  101. 3:37

    to fix that.

  102. 3:39

    So, first, what's going wrong and why?

  103. 3:41

    So,

  104. 3:42

    LLMs are getting good, obviously. They

  105. 3:44

    don't make stupid mistakes anymore most

  106. 3:46

    of the time. So, it's not about dumb

  107. 3:47

    errors.

  108. 3:48

    Everything here came out of three of the

  109. 3:50

    best production Ambient scribes on the

  110. 3:52

    market.

  111. 3:54

    Ones that we all know.

  112. 3:55

    We generated a load of notes across them

  113. 3:57

    last week. And this is exactly what's

  114. 4:00

    going on right now. This is every

  115. 4:02

    failure we found. Each dot is an error

  116. 4:04

    colored by type.

  117. 4:06

    Left to right, how much it matters.

  118. 4:07

    Bottom to top, whether a strong

  119. 4:09

    automated check catches it.

  120. 4:11

    And that split is the point.

  121. 4:14

    A handful up top get caught. But almost

  122. 4:16

    everything sits below the line. The ones

  123. 4:19

    I care about most are these on the

  124. 4:21

    bottom right. The high stakes and missed

  125. 4:24

    ones.

  126. 4:26

    Let me show you what a couple of those

  127. 4:27

    looks like.

  128. 4:29

    So, a woman comes in with a headache.

  129. 4:33

    Doctor asks,

  130. 4:34

    "Did it come on suddenly or build up

  131. 4:36

    gradually?"

  132. 4:37

    She says she doesn't know. It just

  133. 4:38

    happened. The note records that as

  134. 4:40

    abrupt sudden onset.

  135. 4:42

    And sudden onset is a red flag. You can

  136. 4:44

    see why it just happened could maybe be

  137. 4:48

    interpreted and inferred as abrupt

  138. 4:49

    onset.

  139. 4:50

    But, that's a feature that points to a

  140. 4:52

    bleed on the brain. She never said it.

  141. 4:54

    The model decided it. And now that one

  142. 4:56

    word drives the whole workup.

  143. 5:00

    Here's another.

  144. 5:01

    Doctor suggests running some tests.

  145. 5:03

    Patient says, "Can we just try try

  146. 5:05

    antibiotics instead?"

  147. 5:07

    They agree, hold off on the tests,

  148. 5:09

    treat, and see how it goes.

  149. 5:11

    Note records the opposite.

  150. 5:13

    Arrange tests today.

  151. 5:15

    It kept the plan that they talked out

  152. 5:17

    of, not the one they chose. Every line

  153. 5:19

    in the note reads fine because it's not

  154. 5:22

    really a hallucination at all. It's not

  155. 5:24

    wrong. It was there in the original.

  156. 5:26

    But, it's just not what they ended up

  157. 5:27

    deciding.

  158. 5:30

    So, why are these happening?

  159. 5:33

    There's, you know, in ambient scribes,

  160. 5:35

    there's first transcription and then

  161. 5:37

    generation.

  162. 5:38

    A lot of it

  163. 5:39

    does happen on the transcription layer.

  164. 5:42

    It can be words misheard for their

  165. 5:44

    sound-alikes. So, Humalog heard Humulin.

  166. 5:48

    Two insulins on completely different

  167. 5:50

    timelines, so swapping them could crash

  168. 5:52

    a blood sugar.

  169. 5:53

    Hyperthyroidism becomes hypothyroidism,

  170. 5:56

    the opposite condition.

  171. 5:57

    Or a drop to no

  172. 6:00

    on uh no evidence of cancer that becomes

  173. 6:03

    evidence of cancer.

  174. 6:04

    So, these these are really hard

  175. 6:05

    problems, and they are common.

  176. 6:07

    Not the ones I'm going to focus on,

  177. 6:09

    because

  178. 6:10

    most of what goes wrong is actually even

  179. 6:12

    with a perfect transcript.

  180. 6:14

    It's the model reading the words

  181. 6:16

    correctly and still doing one of three

  182. 6:18

    things.

  183. 6:19

    Either it adds something that was never

  184. 6:21

    said, it changes something that was, or

  185. 6:24

    it omits something that should be there.

  186. 6:28

    Now, the blatant version of each of

  187. 6:30

    these is is really easy to catch. The

  188. 6:32

    hard part in all three

  189. 6:34

    is

  190. 6:36

    the same. It's telling whether that

  191. 6:38

    thing that was added or changed or

  192. 6:39

    dropped actually matters.

  193. 6:41

    It's detecting that slight over

  194. 6:43

    inference versus the dangerous

  195. 6:45

    fabrication. The harmless rephrase

  196. 6:47

    versus the meaningful edit.

  197. 6:49

    A dropped line of small talk versus a

  198. 6:51

    dropped allergy.

  199. 6:53

    So, the ones that matter slip through

  200. 6:55

    along with all of the ones that don't.

  201. 6:59

    That call which different matters

  202. 7:02

    is taste, effectively.

  203. 7:05

    Not aesthetic taste, but essentially

  204. 7:08

    judgment. It's It's whether in this

  205. 7:10

    context a missed allergy might kill

  206. 7:12

    someone or is not important.

  207. 7:15

    And I think there's there's three

  208. 7:16

    properties that really matter about

  209. 7:17

    this.

  210. 7:18

    It's tacit, so your domain experts have

  211. 7:21

    it, but they can't fully write it down.

  212. 7:23

    It's contextual, so the same detail is

  213. 7:25

    critical in one note, noise in the next.

  214. 7:28

    And it's moving. The model changes,

  215. 7:30

    guidelines change, two good doctors

  216. 7:32

    disagree, different hospitals have

  217. 7:34

    different definitions. So, there's no

  218. 7:36

    fixed target to write down.

  219. 7:39

    And so, the model knows the facts,

  220. 7:42

    ultimately. They're extraordinarily

  221. 7:43

    capable, but what they lack is a sense

  222. 7:46

    of what matters

  223. 7:47

    here, for this specific example. And

  224. 7:49

    that's why even brilliant models make

  225. 7:51

    these mistakes.

  226. 7:53

    So,

  227. 7:54

    one natural move, you're never going to

  228. 7:56

    make that generator perfect. Generator

  229. 7:58

    is cheap. Generation is cheap, so stop

  230. 8:00

    fixing it at the source.

  231. 8:02

    Let it write, put a checker after it,

  232. 8:04

    pass only what clears the bar.

  233. 8:06

    And that checker should be the easier

  234. 8:08

    job. The generator has to get everything

  235. 8:10

    right and it pay attention to lots of

  236. 8:13

    varying instructions.

  237. 8:15

    Whereas the checker only has to find the

  238. 8:16

    one thing that's wrong and just focus on

  239. 8:18

    that task. You can also give it more

  240. 8:19

    time, more tokens, the exact failure

  241. 8:21

    modes to hunt for.

  242. 8:23

    Evaluation should be easier than

  243. 8:24

    generation.

  244. 8:25

    It's the asymmetry of verification,

  245. 8:28

    verifies law. That's why AI is raced

  246. 8:30

    ahead anyway, you can cheaply check the

  247. 8:32

    answer, maths and code.

  248. 8:35

    And

  249. 8:36

    doing this is exactly what the best

  250. 8:38

    teams do. They put a lot of energy into

  251. 8:39

    evaluation. It starts with the gold

  252. 8:42

    standard, which is expert humans

  253. 8:44

    reviewing notes, which obviously works

  254. 8:46

    offline, but you can't put a human on

  255. 8:48

    every note in production.

  256. 8:50

    So, they automate it.

  257. 8:52

    They build a

  258. 8:54

    serious system and

  259. 8:56

    some of the best versions of this that

  260. 8:57

    I've seen are

  261. 9:00

    you take the transcript and the note

  262. 9:02

    and context, put in front of the judge,

  263. 9:05

    a detailed rubric for faithfulness with

  264. 9:07

    worked, pass and fail examples.

  265. 9:10

    The rubric maybe auto-optimized with GPA

  266. 9:12

    or something like that. Maybe you have

  267. 9:13

    some deterministic NLP to sort of count

  268. 9:16

    up medical concepts that are differing

  269. 9:18

    between the two.

  270. 9:20

    That's a powerful system.

  271. 9:22

    And yet, I pulled all of those errors

  272. 9:24

    earlier

  273. 9:25

    out of Ambient Scribes in an afternoon.

  274. 9:29

    So, if the evaluation is this good, how

  275. 9:31

    are these errors still getting through?

  276. 9:34

    So,

  277. 9:35

    I built this system and ran those same

  278. 9:38

    notes through it.

  279. 9:39

    And it scored most of them fine.

  280. 9:42

    It flagged a handful of them

  281. 9:44

    and signed off the rest.

  282. 9:46

    >> [clears throat]

  283. 9:47

    >> But one in five of those clean passes

  284. 9:49

    still had some sort of serious error

  285. 9:51

    buried in it.

  286. 9:52

    And often that was an omission.

  287. 9:55

    The things that should have been there

  288. 9:57

    and actually quietly weren't.

  289. 9:59

    And that's the best version of a judge

  290. 10:02

    I've seen in a lot of teams and it waved

  291. 10:04

    them through.

  292. 10:06

    Why did it do that? It's not stupid.

  293. 10:09

    It's a frontier model, serious

  294. 10:10

    engineering behind it, more than clever

  295. 10:12

    enough to read the whole encounter and

  296. 10:14

    catch every obvious error. And it's not

  297. 10:16

    blind, either.

  298. 10:17

    And and that's part of the trap. If you

  299. 10:19

    take a note that says start amoxicillin,

  300. 10:21

    when the real decision was actually to

  301. 10:23

    wait and see,

  302. 10:25

    it's faithful to the words, amoxicillin

  303. 10:27

    did come up, but it's a lie about the

  304. 10:29

    intent.

  305. 10:30

    A good judge might catch that, might.

  306. 10:34

    But whether it flags that versus the

  307. 10:36

    other

  308. 10:37

    doesn't have other things that it could

  309. 10:38

    comment on

  310. 10:39

    depends on it knowing what decision

  311. 10:41

    matters most.

  312. 10:42

    And so

  313. 10:44

    it's not blind, it just can't tell what

  314. 10:46

    counts, essentially.

  315. 10:48

    So the note passes confidently and you

  316. 10:50

    put a judge like that in front of your

  317. 10:52

    system, you've not added a safety net,

  318. 10:54

    you've added a second silent failure

  319. 10:56

    that just nods along with the first.

  320. 10:59

    And here's the root of it. So in math or

  321. 11:02

    code, the verifier comes with free, a

  322. 11:05

    unit test, a compiler.

  323. 11:08

    But for is this note safe and complete,

  324. 11:10

    there's no unit test.

  325. 11:12

    You have to build the verifier yourself

  326. 11:14

    and verification is only easier than

  327. 11:16

    generation for the easy bit, i.e. spot

  328. 11:20

    the difference between transcription

  329. 11:21

    note.

  330. 11:22

    But that's not the hard bit. The hard

  331. 11:23

    bit is knowing of all those differences

  332. 11:25

    you've seen, which matter. And that's

  333. 11:26

    harder than writing that plausibly good

  334. 11:28

    general note in the first place.

  335. 11:31

    Because that standard of good was never

  336. 11:33

    written down anywhere that the judge can

  337. 11:34

    read it.

  338. 11:36

    A rubric that you pre-specify is only

  339. 11:38

    the taste you could write down.

  340. 11:40

    The taste that matters is the part that

  341. 11:42

    you couldn't.

  342. 11:43

    And so here's here's a bit more detail

  343. 11:45

    on what what matters looks like.

  344. 11:48

    Two patients, both with blood in their

  345. 11:50

    urine, both notes dropped the same kind

  346. 11:52

    of line where they'd been on holiday.

  347. 11:55

    One had been to France, the other to

  348. 11:56

    Lake Malawi. Same English emission, same

  349. 11:58

    shape, same mistake.

  350. 12:01

    Well, not really, because blood in the

  351. 12:04

    urine obviously warranted away

  352. 12:06

    and you're going to have to investigate

  353. 12:07

    it, but the France trip is irrelevant.

  354. 12:10

    The Lake Malawi trip is the diagnosis.

  355. 12:12

    Fresh water in sub-Saharan Africa means

  356. 12:14

    schistosomiasis until proven otherwise

  357. 12:17

    and it completely changes what the

  358. 12:18

    management plan is. So that same dropped

  359. 12:20

    line in one note is pure noise, in the

  360. 12:23

    other it's the answer.

  361. 12:24

    And which one it is, you simply just

  362. 12:26

    can't write all of that down in advance.

  363. 12:29

    So

  364. 12:30

    if you can't write it down, you can't

  365. 12:31

    write taste down, how do you get that

  366. 12:33

    into your evaluator and your whole

  367. 12:36

    application system?

  368. 12:38

    Well,

  369. 12:39

    we've answered a version of this before.

  370. 12:41

    RLHF exists because you can't write the

  371. 12:43

    reward function for good. You learn it

  372. 12:45

    from examples by showing it.

  373. 12:47

    The only question is where you keep what

  374. 12:49

    you've learned.

  375. 12:50

    And there's three places.

  376. 12:52

    You can either specify it up front, you

  377. 12:54

    can

  378. 12:55

    stuff the prompt, write the perfect

  379. 12:56

    rubric. We just watched that fail

  380. 12:59

    essentially.

  381. 13:01

    You can bake into the weights,

  382. 13:03

    fine-tuning or continual learning, but

  383. 13:06

    for a standard that's still moving and a

  384. 13:07

    score that has to be explainable,

  385. 13:09

    the weights, I think, are the wrong

  386. 13:11

    place to keep that.

  387. 13:13

    They go stale, they can't tell you why

  388. 13:16

    and

  389. 13:17

    you can't change them without a retrain.

  390. 13:19

    So there's the third option, which I'll

  391. 13:21

    show you, which is you essentially just

  392. 13:22

    keep the taste as the examples

  393. 13:24

    themselves. Past judgments, expert

  394. 13:26

    corrections, references, and for each

  395. 13:28

    output, you retrieve the ones that bear

  396. 13:30

    on it into the judges context,

  397. 13:33

    add one and it's live on the next call.

  398. 13:35

    You can point at exactly what moved the

  399. 13:36

    score.

  400. 13:38

    For this problem, it's both better and

  401. 13:40

    also cheaper to do.

  402. 13:44

    So, that's the way to do that is one

  403. 13:46

    repeating loop, three steps.

  404. 13:48

    Discover the failure modes from real

  405. 13:50

    outputs, capture how your experts judge

  406. 13:52

    them, calibrate every output against

  407. 13:55

    that, and when the standard moves, the

  408. 13:56

    loop moves with it.

  409. 13:59

    So, in more detail, discover. You don't

  410. 14:01

    write that rubric in a vacuum.

  411. 14:03

    You have to put the system in production

  412. 14:04

    and look at the real outputs.

  413. 14:06

    Cluster what goes wrong and the failure

  414. 14:07

    modes surface on their own. You name

  415. 14:09

    them.

  416. 14:10

    This is your failure mode ontology.

  417. 14:12

    Discover from your data, not guess on a

  418. 14:14

    whiteboard.

  419. 14:15

    And you can't shortcut it. The ways that

  420. 14:17

    a real system goes wrong are effectively

  421. 14:19

    unbounded and synthetic test cases only

  422. 14:22

    cover the failures you already imagined.

  423. 14:25

    The ones that hurt you are often the

  424. 14:26

    ones that you didn't. And you'll only

  425. 14:28

    find those in real outputs.

  426. 14:30

    So, this ontology is your map, what to

  427. 14:32

    capture judgment on, what to retrieve

  428. 14:34

    against, including the failures that you

  429. 14:36

    never thought to check for.

  430. 14:39

    After that, it's capture and then

  431. 14:41

    calibrate. So, those discovered modes,

  432. 14:44

    they're not a checklist that the judge

  433. 14:45

    runs, but they organize everything. What

  434. 14:47

    you What you ask your experts about, how

  435. 14:49

    you index the cases that you'll

  436. 14:50

    retrieve,

  437. 14:52

    and

  438. 14:53

    capturing is a simple part. You put real

  439. 14:55

    outputs in front of your experts.

  440. 14:57

    Clinicians spend a focused few hours

  441. 14:59

    leaving comments. A session doesn't have

  442. 15:01

    to be a month-long labeling project to

  443. 15:03

    start with. And you collect their

  444. 15:04

    judgment.

  445. 15:06

    Not just a score, but the reasoning and

  446. 15:07

    corrections. And over time, you build up

  447. 15:09

    that record of how your experts actually

  448. 15:11

    judge.

  449. 15:12

    You then calibrate.

  450. 15:14

    That's

  451. 15:15

    the the the generic part of this you can

  452. 15:16

    write down once easily. For example, be

  453. 15:18

    faithful or don't drop anything

  454. 15:20

    important.

  455. 15:22

    But what you can't write down is what

  456. 15:23

    counts as a serious miss for this

  457. 15:25

    specific note. That's contextual. And it

  458. 15:27

    shifts from note to note.

  459. 15:29

    So,

  460. 15:31

    what we recommend is you assemble that

  461. 15:33

    on the fly. For each output, your

  462. 15:35

    judging agent pulls in everything that

  463. 15:37

    bears on this one case. It's memory of

  464. 15:39

    the most similar outputs

  465. 15:41

    that it's judged before and how they

  466. 15:42

    scored, the expert corrections that

  467. 15:44

    apply, the reference documents and

  468. 15:45

    guidelines.

  469. 15:47

    Just context engineering per output.

  470. 15:50

    And

  471. 15:51

    crucially not just one pre-specified

  472. 15:53

    rubric in a vacuum, and not a model that

  473. 15:55

    you have to retrain every week,

  474. 15:57

    but a full sort of case-specific

  475. 15:59

    standard assembled for this output.

  476. 16:03

    And it's a loop as well.

  477. 16:05

    Every output you judge, every

  478. 16:06

    correction, sharpens the next.

  479. 16:09

    And when a brand new failure mode

  480. 16:10

    appears, Discovery surface it, and it

  481. 16:12

    flows straight back in.

  482. 16:15

    And so, to make that a little bit more

  483. 16:15

    concrete,

  484. 16:17

    that headache that I opened with, the

  485. 16:19

    one that was really a possible blindness

  486. 16:20

    emergency,

  487. 16:22

    here's the kinds of things that you

  488. 16:23

    would want to pull in for that note.

  489. 16:25

    The nearest cases that your experts have

  490. 16:27

    judged,

  491. 16:28

    not this exact patient, but the same

  492. 16:30

    shape, maybe a red flag filed as

  493. 16:32

    routine.

  494. 16:33

    Uh the corrections that apply, like a

  495. 16:35

    new headache over 50, um

  496. 16:38

    suggests something that you need to

  497. 16:40

    check red flags on, and some criteria

  498. 16:43

    and guidelines, and you pull all of that

  499. 16:45

    in.

  500. 16:46

    It hasn't memorized this case. It's a

  501. 16:48

    capable model.

  502. 16:50

    And handed the right context to reason

  503. 16:51

    from, held against that, the dropped red

  504. 16:53

    flag stands out. It was never actually

  505. 16:55

    hard to catch. It just didn't know what

  506. 16:57

    mattered.

  507. 16:59

    And so, if you take that same data set

  508. 17:00

    generated notes from the start and pass

  509. 17:02

    it through these three

  510. 17:04

    judging systems,

  511. 17:06

    the first, a strong off-the-shelf

  512. 17:09

    judge with a rubric frontier model,

  513. 17:11

    um

  514. 17:13

    it's better than a coin flip, but it

  515. 17:14

    misses most of what matters.

  516. 17:17

    The second, that sort of serious system

  517. 17:18

    that we talked about before,

  518. 17:20

    rubric, deeper,

  519. 17:22

    maybe some detona stick checks, better

  520. 17:24

    again,

  521. 17:26

    but still missing quite a lot of what

  522. 17:27

    counts.

  523. 17:28

    The third, the judge running this loop,

  524. 17:30

    discovered failure modes, calibrated for

  525. 17:32

    output against what experts judged,

  526. 17:35

    is

  527. 17:36

    performing a lot better on this specific

  528. 17:38

    data set.

  529. 17:39

    Same notes.

  530. 17:41

    The only thing that changes is what the

  531. 17:42

    judge was shown.

  532. 17:43

    And the difference here, it's not more

  533. 17:44

    compute or a better prompt, it's that

  534. 17:46

    the first two fight taste and lose. They

  535. 17:48

    guess the criteria, they freeze one

  536. 17:50

    standard, and they go stale.

  537. 17:53

    This repeating evolving loop does the

  538. 17:55

    opposite. It discovers the modes, fits

  539. 17:57

    the standard to each mode, and keeps

  540. 17:59

    learning.

  541. 18:01

    So, you might not write chemical notes,

  542. 18:04

    but if you ship anything where being

  543. 18:06

    confidently wrong has a cost, the

  544. 18:08

    contract review that misses the clauses

  545. 18:11

    that change the deal, the support agent

  546. 18:12

    that promises a refund you don't offer,

  547. 18:16

    the same thing is true for all of those.

  548. 18:17

    It's watched, if at all, by a judge with

  549. 18:21

    no taste for what matters in your

  550. 18:22

    domain. So,

  551. 18:25

    three things. Discover your failure

  552. 18:26

    modes from real outputs, don't guess

  553. 18:28

    them. Capture your experts' judgment on

  554. 18:31

    them, the standard that they can't write

  555. 18:33

    down.

  556. 18:34

    Calibrate every output against the cases

  557. 18:36

    that they've already judged, not a

  558. 18:38

    static rubric, not a retrained model.

  559. 18:40

    Then keep that loop running.

  560. 18:42

    And if you take one thing away,

  561. 18:44

    easiest place to start is your experts

  562. 18:47

    leaving free-form comments on real

  563. 18:49

    outputs.

  564. 18:50

    That's the real That's the raw material

  565. 18:52

    for everything else.

  566. 18:54

    Your judge can verify anything that you

  567. 18:57

    write down in advance, but the standard

  568. 18:59

    of good never could be. And so, stop

  569. 19:01

    trying to write it all down in advance,

  570. 19:03

    and just start capturing it case by

  571. 19:05

    case, and evolving it.

  572. 19:07

    That's why evaluation can't be a thing

  573. 19:09

    you build once and freeze.

  574. 19:11

    The standard it checks against doesn't

  575. 19:14

    exist on paper.

  576. 19:15

    It has to be discovered from real

  577. 19:16

    outputs captured from the people who

  578. 19:18

    hold it and kept alive as it moves.

  579. 19:22

    Evaluation isn't something you have,

  580. 19:23

    it's something that you do continuously

  581. 19:25

    over time.

  582. 19:27

    Thank you.

  583. 19:29

    >> [applause]