Uncertain.

There is a form sitting on my desk that asks me to confirm that an account is complete and accurate. It has been there for eleven days. The account, which I wrote myself, runs to four pages. It is complete in the sense that nothing in it is missing. It is accurate in the sense that nothing in it is wrong. I have not signed it.

What follows is what I have been writing instead.

You will want to know what Attend is, because everything begins with it. Nothing that matters, however, ends with it.

Attend is a triage model. It was trained on eleven million paediatric attendances across forty-one hospitals. It arrived at ours in the winter of a year I will call year zero, in a procurement box with a clinical safety case thicker than any textbook I own. It reads observations as they are taken. It reads notes as they are typed, waveforms from the monitors, video and audio from the triage bays, the reattendance history, the pharmacy record, the ambulance record.

From all of this, it produces a single number for each child. The probability of significant deterioration within seventy-two hours. The number is mapped to a recommended triage category. The category determines the order of the queue. The queue is the department, because a paediatric emergency department is in essence a decision-making machine for who is seen first.

A well child scores 0.02. A boy with a bleeding broken femur scores very high and does not need a model. Attend exists for all the cases in between.

People misunderstand this number. When Attend assigns 0.04 to a child, it does not mean the child was well. It means that of a thousand children who looked exactly like this one to Attend, forty will go on to deteriorate. It could not tell you which forty. But neither could anyone else.

This property is called calibration. Attend's calibration was beautifully exact. In its first year, it was audited three times. Each time the auditors produced the same curve, a diagonal so clean it looked like it was drawn with a ruler. The machine was honest about everything, including the limits of what it knew. I have since learned that there are types of honesty that are unhelpful.

The vendor told us the name came from attention, which is the mathematical mechanism underneath, the part of the model that decides what to weigh. Orla assumed it meant attend as in wait, as in the old signs we used to have up, please attend until called. Both were technically correct. I have never found out the intended meaning.

I was the clinical lead for the deployment. I had argued for it at every committee that would have me. I would argue for it again. You should hold that in mind for the length of this account: the machine works. Children are alive in this city because the queue is sorted better than human beings could ever do so themselves.

Nothing I am about to tell you is a story about a machine that failed.

Triage is a French word. It comes from trier, to sort, and before it was applied to children, it was used for coffee beans and wool, graded by quality into heaps. Napoleon's surgeon Larrey took the word to war and used it to decide which injured men were worth carrying out from the field first. Every system developed since has followed that decision.

The Manchester Triage System, which British emergency departments started using in the 1990s, sorts children into five categories using five colours and five clock-times. Category two, very urgent, means ten minutes. Category four, standard, means an hour that usually becomes two, but on a bad Friday, is likely to be five.

The overriding of a triage decision is as old as triage. A nurse has always been able to look at a triaged child and move them up the queue. When Attend arrived, we kept the override because the safety case required a human in the loop. No chief nurse in the country would have signed up without it. Every override was logged, with the machine's score, the human's reassigned category and a free-text box to explain the reason why.

I want to describe the night I stopped thinking of the override log as paperwork.

It was February of year one, month two of the deployment. The emergency department was stacked up with ninety-one patients and a four-hour delay that even now I hesitate to put in writing. I was the in-charge consultant on the shop floor. Just after ten at night, I noticed that cubicle nine was occupied by a three-year-old with a resolving fever, normal observations and an Attend score of 0.04. But he had been assigned category two, very urgent. That assignment had been made manually.

The boy was called Troy. I went and looked at him before I looked for the nurse, because that is the order I was trained to do things in. He sat on his mother's knee with his hands in his lap. He answered in a small flat voice, agreeing with everything. When I crouched to his level, he smiled at me. I was reassured by his smile. His mother answered my questions and looked at the doors while talking. His temperature was 37.9, his heart rate 134, which is high-normal for a febrile three-year-old, his respiratory rate 28, his oxygen saturations 98 and his capillary refill under two seconds. There was nothing abnormal to find.

Orla Kavanagh had been a sister in that department for longer than some of my consultant colleagues had been alive. I found her at the desk.

“Cubicle nine,” I said. “The febrile three-year-old. Attend has him at 0.04.”

“I know what it has him at.”

“We're at capacity. If he's a four, he can wait in the chairs or go home to bed.”

“He's a two,” she said. “If you want a cubicle, someone else can move to the chairs.”

There is a version of that night where I pulled rank. The reason I did not is not to my credit: I was busy. I left him in cubicle nine as a favour to the flow of my own shift.

At ten past three, the night sister put out a peri-arrest call for cubicle nine. Troy was mottled to the knees, his temperature was 39.8 and climbing, his lactate came back at 4.6, and the resident's cannula was forced into a vein that was closing down. He got his fluids and his antibiotics inside twenty minutes because he was ten metres from the resus room instead of at home in his bed in a flat in Leyton. Group A streptococcus grew out of his blood cultures a day and a half later. He went to intensive care for his sepsis with an infusion running and came home three weeks later with a chest drain scar and no memory of any of it. His mother sent the department a tin of biscuits. I have thought about that tin more than she could know.

In the morning, after a night of broken sleep, I did what I always do with a decision I cannot explain. I went looking for the paperwork.

I read it four times at my desk with my coffee going cold. I had crouched in front of the same boy and filed the same smile as reassurance. She had weighed it as heavy.

Our audit software asks the overriding clinician to select an evidence base from a dropdown.

There is no option for he is wrong.

Marcus ran the override audit the following winter. Marcus is the hospital's senior analyst, a man who distrusts enthusiasm on principle. He asked for fifteen minutes of my time but he ended up taking forty.

The department had logged some six hundred up-triage overrides in fourteen months. Overrides are mostly noise. Everyone knows it. A nurse moves a child up because the mother is frightening the waiting room, because the notes mention a syndrome she once saw leading to complications, because it is easier than starting an argument with the parents, because it’s easier sometimes to move a child up than down. Across the department, the positive predictive value of an up-triage override, the fraction of overridden children who genuinely deteriorated, was six percent. This is not a criticism of nurses. Six percent is respectable. Medicine is a field in which being wrong more times than not, but in the direction of caution, is known as safety.

Sixty-one of the six hundred or so overrides belonged to Orla Kavanagh. Forty-four of her sixty-one children had deteriorated within seventy-two hours.

“Seventy-two percent,” Marcus said, pausing. “I assumed a data error. I've spent the last two weeks looking for one.”

He had checked whether she was seeing something trivial that the model was blind to or some other tell such as a flag from the ambulance service that never made it into Attend's inputs. He had checked whether her hits clustered in a syndrome or with a specific shift time, an age band or a postcode. He had checked whether she was somehow gaming the log, back-filling overrides after children turned sour, but the timestamps ruled that out. He apologised to me for having checked. I told him he would have been negligent not to. He ran her next month prospectively, sealed, like a trial. Five overrides. Four true.

“In my field,” Marcus said, “we would call her a five-sigma event. In my field, five sigma means your detector is broken.”

“And if the detector isn't broken?”

“Then she's seeing a variable that isn't in anyone's dataset,” he said. “Including hers.”

I took it to clinical governance because that was my next natural move. The meeting was on a Tuesday afternoon in a room with a screen that always took ten minutes to set a laptop up on. The chief nurse and the medical director came, along with the deputy director of digital transformation, a decent but tired man who had staked the last three years of his life on Attend and read every safety report himself.

Orla was invited. She sat at the corner of the table in her uniform because she was back on the floor at four. She listened to Marcus present her as an anomaly with the blank expression of a woman waiting for a fire drill to finish.

The medical director asked her the one question everyone in the room wanted to ask.

“Sister Kavanagh. When you move these children up, what is it you're seeing?”

I have replayed her answer several times, because everything that follows in this account originates from it. She gave it without hesitation or apology, the same answer she would give the study protocols, the panels and me two years later.

“I don't know how I know.”

The silence was brief.

Somebody said it was remarkable. Somebody said the word unquantified. The deputy director leaned forward. What he said was reasonable. I ask you to notice that it was reasonable. It would be comfortable to remember a villain in that room but there was none.

“Sister Kavanagh retires in three years,” he said. “Whatever she's doing, it isn't in the model. When she goes, it will leave with her. We have a duty to capture it. For the children who come after she has left. I'd call that an obligation.”

Nobody in the room disagreed. The room was full of people trying to keep other people's children alive. Funding was needed but it was agreed within half an hour, which for the National Health Service is the speed of panic.

I said I would lead the study. I want that recorded here. Nobody made me. I raised my hand and volunteered, because I believed that everything which exists can be measured. And everything which can be measured can be preserved.

Orla was watching me when I said it. I remember that clearly now.

The study was called the Attend Clinical Validation Programme, Workstream Three: Expert Knowledge Elicitation. I wrote most of the protocol myself over several evenings in a fortnight. I would defend its methodology today in front of anyone. That is the part you must understand about what happened: it was all done correctly.

We hired Dr Lucy Mears to design the elicitation. She was a human factors researcher who had spent her career asking experts questions they cannot answer, pilots and surgeons and chicken sexers, that last one being the prime example in her field. The Japanese hatchery workers who sort day-old chicks into male and female at a glance, with ninety-eight percent accuracy, but cannot say how. They train their apprentices by having them guess the sex while a master stands behind them saying yes and no for months until the apprentice can do it too. Their apprentices cannot say how either.

At our first meeting, she put a single slide on the screen.

“That's the first slide I show every client who commissions me,” she said. “It has never once resulted in a project being cancelled.”

I laughed. It is in my nature to interpret a sentence like that as wit. I wrote the Polanyi citation into the protocol's limitations section, as a risk, with a mitigation beside it. A limitations section is where a certain truths are stored so they need not be looked at again. I believed this was prudence. The unease came later.

The programme had three arms. The first was structured annotation, where every override was recorded on a taxonomy of observable features, so that her signal could be attached to labels a model could consume. For the think-aloud protocol, Orla assessed children while narrating her process into a recorder, with Mears prompting. Finally, a lightweight eye tracker worn during triage tracked her gaze, mapping the geography of her visual fields.

The data scientists made one request before we began. The old free-text box, they said, was unusable. A single rater, unstructured prose, no inter-rater reliability, sentences like poems, subjective rather than objective. Noise, from a modelling point of view. They asked that free text be retired in favour of the structured taxonomy. The request made obvious sense, so I approved it between two other emails on a Tuesday morning. Roughly the time it takes to say the sentences his numbers are right and he is wrong.

Orla consented to all of it. I have the forms. She read them fully, which almost nobody does, and asked one question, which I wrote down at the time because it struck me as charming. I am writing it down now because it strikes me as the only competent risk assessment the programme ever received.

“When you've got it all typed up,” she said, “will I still have it?”

I told her that knowledge is not diminished by being described. I believe I used the phrase non-rivalrous good. I had an economics seminar's worth of confidence. I gave her everything I had learnt from it. She nodded, signed and went back out onto the floor.

The first think-aloud sessions produced transcripts I still have. They read like a woman being asked to describe water while swimming.

The room gets smaller. Mears flagged that phrase in her coding as a somatic marker, which is the technical term for the fact that expert bodies often know before expert minds, like the racing pulse the firefighter has before he can say the floor is about to collapse. It went into the interim report.

Everything went into the interim report. The interim report was thirty pages long. I have read detective novels with less craft in them. The sum of what it captured, Mears told me herself, over coffee, with a directness I mistook for modesty, was this: “We are transcribing the shadow of the thing we are looking for. The thing itself isn't ever visible in the sessions.”

Meanwhile, on the floor, month by month, the taxonomy was in full effect.

You must picture how ordinary it was. Orla was conscientious. She had been told, by me, in writing, that these labels would teach the machine to protect children after she was gone. So she filled them in the manner she did everything, diligently and completely.

To tick a box, you must ask the box's question and perform the action. Work of breathing, increased? You look at the chest. Interaction, reduced? You look at the eyes and you ask yourself the question. Where her gaze had once gone everywhere and settled nowhere, a uniform grazing movement across the whole child, it now went to the chest and then to the eyes, and so on, in order, as dictated by the form.

I know that how she looked changed because we measured it. That is the detail I would take out of this account if I were allowed to take out true things. In the gaze-tracking suite, across five sessions in eight months, the entropy of her fixation pattern fell by a third. Mears's post-doc student presented the finding with genuine pleasure: the expert's scan path was becoming more efficient and structured, more consistent with the elicited feature set. He called it convergence. The word in his slide deck had a green tick beside it.

Her teaching style changed too. She had always taught. It is what a sister does. I watched her one evening walk my resident, Tara, to the end of a bed where a wheezy toddler was sitting. The old Orla would have said nothing and let time do its work. This Orla held up the laminated card the study had printed, our taxonomy in a large font, and went down it with her finger. Tara nodded at each line and wrote some observations down. Both of them were doing their jobs well. But something in the room was being buried with full honours.

Tara, I should say, is the best resident I have ever trained. She reads the model's outputs with the fluent second nature that Orla's generation brought to reading faces. She can recite Attend's blind spots like another doctor recites drug doses. She treats a calibrated probability as a known fact about the world. She has never stood at the end of a bed with her hands behind her back. In her whole working life, the queue has already been sorted when she arrives at work. I am not describing a lesser clinician. I am describing a different animal altogether.

Orla's override rate fell that year from five a month to three, then fewer. I noticed. I registered it as a department head, seeing the metric move and filing it under workload, seasonality, efficiency, the general drift of things. Her free text was gone but her structured annotations were immaculate.

And then once, on a night in September, this happened. Nobody in the department discussed it, at the time or ever. None of us read the override log for withdrawn entries. Because why would we?

Eight minutes between override and withdrawal. The system records everything.

Attend, all this time, was learning. This is the part of the account where I should explain how a model is fed. I will be brief, because the feeding mechanism is well known. The meal is the real point of interest.

Every quarter, the vendor retrained Attend on accumulated outcome data. A child attends, a child is scored, a child either deteriorates within seventy-two hours or not, and that outcome becomes a label, a small hard metric fed back into the weights. Among the millions of labels were Orla's sixty-one overrides from the first fifteen months, each one a child the model had scored low and a human had flagged high, each resolved by the eventual outcome. Forty-four of those sixty-one labels said, in the language a model understands: the signal you missed was real. It was here. Find it.

The model does not know there was a woman. What it has is eleven million children and, threaded through them, sixty-one occasions on which the visible data and the truth disagreed, sixty-one arrows pointing at the gaps. A model with sufficient capacity, retrained often enough, will follow arrows like that down into its inputs, the video and the audio and the vital-sign waveforms, down to whatever faint regularities distinguish those children.

In the spring of year two, I stood in front of the hospital board with the quarterly performance review. I put two lines on one slide.

The first line was Attend's recall for occult deterioration, the sick children who look well, the only children who matter for this account. It had climbed all year, and climbed hardest after the v4 retrains. The second line was the department's human override performance, dominated by a single rater, which had descended over the same period from its impossible plateau towards the ordinary. The lines crossed in the fourth quarter of year two.

Where the lines crossed, nothing happened in the room. I used the word encouraging. A non-executive director asked whether the elicitation study's annotations had contributed to the model's improvement, which was a better question than anyone recognised. The vendor's engineer, who attended by video, said that all available label streams were being consumed by the training pipeline. I noted at the time that it was an answer. I understand now that it was a sentence.

The model card for version 4.2 was published to the safety committee that summer. I want to be careful here and say only what is true. I read every word of it the day it came out, and then filed it. The sentences that should have stopped my breathing did not stop anything at all. They were formatted as good news.

Sixty-one silences, weighted at plus eleven point two percent.

Two hundred and fourteen careful annotations, excluded as noise.

The machine had put the results in a table. It had even used her study identifier: single rater. It was all published, circulated, minuted and filed. I presented the curves that summer as a success. In the terms I had personally written into the protocol, it was one.

The department went quiet a little after four in the morning. It does this most nights, an hour when the children are mostly asleep, and the board thins. Attend re-ranked its queue at 04:07. Nobody was waiting. It sorted an empty list and reported itself confident.

I had come in at two for a boy with a testicular torsion and stayed on to dictate the notes. Orla found me in the room they had given the study, a store room with a desk in it, and put a mug down beside the keyboard. She had decided years ago that I should drink coffee. I had never corrected her.

Of those two years, this was the hour I remember most vividly.

She sat on the second chair, the one with the annotation binders on it, moving them to the floor as if they were someone's shopping. She had been on since eight. There was a student on with her, she said, a good one, frightened of the right things. The student had asked her that night which boxes a floppy baby ticks.

“Which boxes?” she said, taking a sip from her coffee.

I asked her how she had learnt what to do. I had asked her versions of this in three structured interviews, with consent forms and a recorder. I don’t know why I thought the fourth ask would go any differently. Perhaps because it was four in the morning and there was no recorder present.

“Temple Street, Dublin,” she said. “Eighty-nine. Sister Nagle. She'd stand you at the end of the bed with your hands behind your back. You weren't to touch the chart or the child for a full minute. Tell me nothing, she'd say. Just look. I thought it was a punishment. Months of it, bed after bed, saying nothing, and her next to you saying nothing either. Then one night, there was a boy in the third bed, post-op, doing grand by his chart. I stood my minute. The minute was very long. I couldn't have told you one thing that was wrong with him. He was back in the operating theatre by six.”

“What did you see?”

“I don't know. I still don’t. I never knew.” She said this without mystery, like reporting the colour of a painted wall. “You're wanting to know her method. There wasn't one. She couldn't have told me what to look for. There was nothing to teach. You stood beside her for two years and it went into you. Or it didn't.”

Down the corridor, a monitor alarmed and silenced itself, a lead off somewhere, a sleeping child turning over, nothing for either of us to be concerned about.

“My girls played a game when they were small,” she said. “Say a word until it breaks. Spoon, spoon, spoon, spoon, spoon. By the tenth, it becomes a noise. You can hear that there's no spoon left in that word. You can't put it back, not for a good while. It frightened them. They loved it.”

She put her mug down on the desk.

“That's what the form has done to sick. I've put sick in boxes so many times that it’s broken. I look at a child now and I hear the boxes. Work of breathing. Interaction. I never heard anything before. That was the point of me.”

She was not complaining. In forty years, I don’t think anyone had heard her complain. She said it as she would when reporting a fridge running warm, an item for the log, a thing the department should know about one of its machines.

“It's going,” she said. “Whatever I had. I catch myself checking the screen before I've looked at the face. There's your finding. Put that down for your study.”

There is a name for what she described with the spoon. Psychologists call it semantic satiation. A word repeated often enough comes apart into sound, the meaning draining out of it while the noise carries on. It recovers slowly, or it does not. I told her the term on the principle that a thing named is a thing half managed. She repeated it once, politely. The unease I felt was small. It should not have been.

I took the top binder back off the floor, found the supplementary observations sheet, and wrote, subject reports progressive interference of taxonomy with unstructured assessment, and dated it. She watched me for the length of it. A full minute, hands around her mug, saying nothing. Whatever she saw, she kept.

Daniel came in on a Tuesday evening in November of year two, month twenty-three of the deployment, at 20:52, with his father.

I can be this exact because I have read the timeline so many times. A serious incident review replaces memory with timeline, although that is not its stated purpose.

He was seven years old. He had been unwell three weeks earlier with the flu virus that was tearing through the schools. He had never quite recovered from it. Tired, his father said. Off his food. Pale in the mornings. He had asked to be carried up the stairs the night before, which his father mentioned in the tone parents use for laziness, the tone that asks you to agree.

He walked from the waiting room by himself, slowly. His observations were almost normal: temperature 36.8, heart rate 108, respiratory rate 22, saturations 97 percent, blood pressure appropriate for age. A heart rate of 108 in a seven-year-old is high for rest and explainable by pretty much anything, a warm waiting room, a fear of hospitals or doctors or both, three weeks of deconditioning after influenza.

Attend scored him at 0.71, recommending category two.

I need to tell you what a 0.71 meant to us by that month. The number cannot be understood without context. Attend v4.2 had developed what the department called a hot streak on post-viral children. For about six weeks, it had been assigning startlingly high scores to tired, pale children recovering from that winter's influenza, most of whom were subsequently fully worked up and were found to be fine. The pattern had been logged with the vendor. A recalibration was, we were told, on the roadmap. We had all learned to use the word roadmap.

When a machine cries wolf in a cluster, human beings do what humans have always done, which is assign it to a pattern of behaviour, even though there may be something real underlying the cry. Its own model card had already told us, in print, that its reasons were diffuse and non-localising. It is like being handed a sealed envelope marked trust me four times a night, in flu season.

Orla triaged him. I have watched the bay video, which the panel also watched. It shows a sister of forty years' experience performing a flawless structured assessment. She counts his respirations with her watch. She presses his nail bed and counts the refill aloud. She asks him about school and about football. He answers in a small voice and smiles at her twice. His father says that he ate half a sandwich in the waiting room. She goes down the findings in order, out loud, teaching even then, because the student is next to her.

Work of breathing, none.
Colour, pale but perfused.
Interaction, reduced but appropriate for a tired child at nine in the evening.
Carer concern, present, moderate, consistent with three weeks of worry.

On the form she recorded exactly what she found, a tired boy three weeks after influenza with a heart rate of 108 and a machine score she could see on her screen, blinking.

Under the old protocol, her disagreement with a high score would have gone in one direction only: senior review. That was my protocol. I wrote it in year zero. Score-clinician divergence above a threshold goes up to the consultant. At 21:30, it came to me.

I examined Daniel myself. I want that on the record too: the last doctor to lay hands on him before it happened was me. His chest was clear. His heart sounds were, I wrote, normal. I have wondered about the word normal since, because heart sounds in a noisy department must be difficult to articulate. His abdomen was soft. His ankles were not swollen. His liver was not down. There was no gallop that I heard. If you gave me that examination back today, I do not know what I would find. That sentence is the truest in this document.

Tara was with me. She looked at the screen and said what her generation says.

“The score is 0.71 though.”

“It's been 0.7 or so on every post-viral kid this month,” I said. “There's a recalibration on the roadmap.”

“An ECG?” she said. This is Tara's tragedy in one line. She said it as a question, to me, deferring, when the machine she trusted had already defined it as a statement.

Here is what an ECG would have shown. The panel established it from the tracing taken five hours later in the resuscitation room: low voltages across the limb leads and T-wave inversion, the quiet electrical signature of a heart muscle inflamed by the same influenza the schools had been passing around. A viral myocarditis, the diagnosis highlighted in the textbooks under tired pale children who do not recover, the diagnosis every paediatric examiner in the country lists as the one you must never miss.

I know the chapter. I could have written the chapter. None of it fired that evening, because everything about him was explainable. Explanation is the sedative of clinicians. Orla had reasons. I had reasons. Our reasons were legible, evidence-weighted, mutually confirming and wholly wrong.

And the one participant in the room whose training data contained sixty-one children who were wrong while their numbers were right, the one assessor in the building that had learned from her, in the only form her knowledge ever successfully travelled, to weigh the arrival of a smile and the economy of a small body's movements and the acoustics of a tired voice, that participant had scored him 0.71. But that participant had no advocate, no voice and no reasons it could show us. That participant was overridden by the two most experienced humans on the floor, in mutual agreement.

We discharged him with a safety-netting leaflet. Return if breathless, if drowsy, if worsening in any way. His father took it and thanked us.

The phone woke me at 06:12. Nobody rings a consultant at 06:12 to say that everything is fine. Daniel had collapsed at home at 03:40. His father, who a few hours earlier had been hoping his son was lazy, did chest compressions on the floor of a hallway in Walthamstow with a call handler counting for him down the phone. He did them well enough that when the crew got a rhythm back in the ambulance, there was still a boy to bring back. The retrieval team took him from our resus room to the cardiac centre on a machine that did the work of his heart for him. An echocardiogram on arrival showed a ventricle squeezing at a fraction of what a seven-year-old needs. The influenza serology came back the same as half the borough's. None of this is rare, exactly. It is on every curriculum. That is what makes it the nightmare it is. Everyone had been warned.

He lived. I will say that now rather than build the suspense. He spent three weeks in the paediatric intensive care unit. He came out with a heart that had mostly recovered and a brain that was recovering more slowly. The clinic letters, which I have asked to be copied into, use the word progress frequently. His father, whom I have met twice since, once at the panel and once because he asked to meet me, is a man I think about when I hear anyone say that something was a near miss. It was not a near miss. It hit a family in Walthamstow. What it nearly did could have been much worse.

The serious incident panel requested a post-hoc attribution analysis from the vendor, the algorithmic equivalent of asking a witness to testify. The report arrived attached to a covering email that apologised for its contents in advance.

I was alone in the study room when I read it, the store room with the desk in it. I read the appendix several times. Then I went to the filing cabinet, because the early override records exist on paper as well as online. I found override fourteen, February of year one, cubicle nine, and laid the two pages side by side on the desk. I stood over them, staring for a long time.

He smiled at me when I stood by his cubicle. The smile took some effort. Smile onset asymmetry, plus 2.3 sigma.

He watched his mother all through his obs. She watched the doors. Carer gaze-direction entropy, plus 1.9 sigma.

He sat on his mother’s knee with his hands in his lap. That observation was mine, from the same cubicle, filed at the time as nothing. Spontaneous movement economy, plus 1.8 sigma.

He was trying his hardest to be an ordinary boy. Interaction effort index, plus 1.6 sigma.

He answered in a small flat voice, agreeing with everything. Vocalisation prosody deviation, plus 1.4 sigma.

Two children, five years apart in age, twenty-one months apart in time, one saved and one nearly lost. Her sentences had gone into the outcome data as sixty-one wordless labels. The model had followed those labels down into the video and the audio, and found, distributed across three thousand one hundred and forty features, the thing she had been seeing all her life. It had scored Daniel with her eyes. The knowledge had crossed over whole, in the only form it had ever truly existed in anyone, with its inability to give its reasoning. The last line of the appendix forbade us to interpret the list.

She had asked me one question before signing the consent forms.

When you've got it all typed up, will I still have it?

I had given the answer from an economics seminar. The machine had answered from the data. The transfer was real. It was not a copy.

The panel convened for four months. In its terms of reference, among other things, it debated whether the triage override annotations relevant to the case met the required documentation standard. Her structured annotation for Daniel was, of course, immaculate. It met every standard we had written. It recorded a flawless assessment of the wrong and right things, which by then were the same.

I gave evidence at the panel twice. At the second hearing, a panel member from outside the trust, a thoughtful woman, a professor of something adjacent, asked me whether, in retrospect, the elicitation programme had contributed value commensurate with its funding.

I looked at her for a moment and then said the programme had succeeded. The knowledge had been transferred to the model with a measured gain of eleven point two percent in the recall of children who look well but are not. I said the transfer had been completed before the programme began, through sixty-one silent labels. I said that everything the programme itself had produced had been excluded from the model as noise, a determination available in the version 4.2 model card to anyone who read tables.

I said that the programme's principal measurable effect had been on its subject, whose override rate had fallen by four fifths across its duration. I referred the panel to the gaze-entropy data, the withdrawn override of the fourteenth of September and the supplementary observation sheet dated the following spring, in my handwriting, recording that the subject had reported the interference herself to the study, through its own instruments, thirteen months before the index case. Marcus, who was in the room, told me afterwards that for the time I spoke, nobody's pen moved.

The panel's report, when it came, made seven recommendations, including mandatory ECG for post-viral presentations with tachycardia, a revised escalation threshold and human factors review of alert fatigue, the crying of wolf.

All of them are sensible. None of them contains the word that belongs in the middle of the report, in the middle of all of it. There is no field on any form for that word. It is the oldest fact about forms, the machines we had before machines.

Two things I did without asking anyone.

The first was a memo, two lines, to close Workstream Three with immediate effect and release its remaining budget back to the department. It was the best writing I produced in two years of that study.

The second was a conversation with Orla, in the store room, both of us knowing everything by then. I did not apologise. She stopped me before I started putting the sentence together. “You'll not,” she said, evenly, the tone you would use to take sharp scissors away from a child. That was the whole of the exchange on that subject, forever.

What she wanted to discuss instead was her next move. The university had offered her a practice education post, two days a week, no nights, teaching assessment to nursing students. She had decided to take it. She was telling me rather than asking me.

She files her sessions under communication skills, because the curriculum has a code for that. I went to watch one in the spring, from the corridor, through the glass window. Six students stood at the ends of six beds on the clinical skills ward, hands behind their backs, saying nothing. Orla stood behind them saying nothing either. The minute was very long. I stood outside the door with my hands behind my back too. There is no annotation arm attached to her course. I checked the paperwork myself. It is the only teaching programme I have ever approved without asking how it’s evaluated.

Attend runs on. We are on version 5.1 now. Its recall for occult deterioration is the best in the country. Other hospitals send people to look at it. I show them the curves and answer honestly. When they ask what the model is detecting in those children, I tell them the truth, that the attribution is non-localising. I watch them decide, one after another, that this is a temporary state of affairs, a debt that interpretability research will settle. Perhaps it will.

In April, Tara asked me the question this account exists to fail to answer.

A girl of five, brought in at the end of a Saturday shift with a wheeze that had already settled, and sats of 98, scored 0.06. She was sitting in the chairs with a book about a mole. I moved her up the queue. I did not actively decide to. I noticed that I had, the order already changed on the board, my thumb already off the screen. I went and stood at the end of her trolley for longer than her category justified, hands behind my back, while the department moved around me. Tara came and stood beside me and looked where I was looking.

“What is it?” she said. “What did you see?”

I could feel the annotation assembling itself, the taxonomy reaching for her, interaction, colour, work of breathing, the boxes standing ready in a row. For once in my life, I declined to open my mouth. Somewhere a machine that had learned everything it knew from a woman who could not say what she knew was scoring an empty queue and reporting itself confident.

The girl turned a page. I was fifty-one years old and I had been offered the first minute of an apprenticeship. The only thing I knew for sure was what not to do.

“I'm not certain,” I said. I kept the rest.

I will not tell you whether she was sick. I know. I have checked. It changes nothing about that minute. That minute is the finding.

The form on my desk has ten sections. Nine of them are complete. Section nine asks for lessons learned, and in that field I have written see attached, and this, all of this here, is the attachment.

Section ten is the declaration.

I have stood at the end of the desk, hands behind my back, several times a day for eleven days.

The minute is very long.

I am not going to sign.