# Chapter 16: Full Scope From *Things That Were Not There*, Book One of *After Intelligence*, by Peter Pink-Howitt (written in collaboration with AI systems; see the Preface). Web page: https://compossible.blog/2026/10/22/chapter-16-after-intelligence-full-scope/ Rights: © 2026 Peter Pink-Howitt. All rights reserved. Free to read, link to and quote briefly with credit; do not republish, translate, adapt or sell the text without written permission. The blog's CC BY licence does not apply to After Intelligence. --- > **Exhibit 16.** Reply to analysis-cell query, M. Lindqvist, 09:07, Day+3. > *Attribution basis is sound. … Model behaviour on this task is within specification.* The query arrived while Lindqvist was still deciding whether the coffee was worth finishing. He had slept four hours on the sofa in his flat with his shoes on, and come in at ten to eight because the building was easier to be in than the flat. The flat had a window that faced the ring road and a refrigerator that held mostly condiments. Every few days, his son in Malmö sent him voice notes about football or school or nothing. Lindqvist played them at double speed on the bus, because at double speed a fifteen-year-old sounded like a fifteen-year-old and not like someone hesitating. There was one from the night before that he had not played yet. It was in his queue with his name on it as owner: the protected status applied at 02:07, and the re-attribution of the 21:31 send. He read the analyst’s note twice, the second time for its grammar. *Rule change of 16:52 expected to raise device to nomination. Nomination suppressed following automated attribution review. Request explanation of attribution basis and of model behaviour.* It was a well-made sentence. It put the expectation first and the model second, so that a reader would arrive at the model already knowing it had done something unexpected, and it asked for an explanation. He opened the attribution review. He had spent eight years, before defence paid better, teaching machines to tell one voice from another for a company in Gothenburg that made hearing aids. He had built speaker models for rooms full of cutlery and cafés with bad acoustics, and had spent his evenings drawing curves that traded one kind of error against the other: the false accept, where the device decided a stranger was the wearer’s wife, and the false reject, where it decided the wearer’s wife was a stranger. He still thought in those curves. Most problems, he believed, were a matter of where you chose to stand on one. The review had stood well. `PRINCIPAL USER: JIBRIL, NADIA. AGE: 11. CONFIDENCE: 0.97.` Voice match across every message sent from the device, against a child heard in two hundred and six recorded lessons. He pulled the match scores and looked at their spread, which was narrow, and at the background noise in the recordings, which was a house: a tap, a pan, someone further away. The match would have passed any standard he had worked to in Gothenburg. Then he looked at the re-attribution of the 21:31 transmission, and the motion data, and the two-minute gap between the London number arriving from her brother’s phone and the clip leaving hers. The accelerometer trace showed a rotation and a short drop, consistent with a phone passing from someone sitting down to someone standing. He had once spent a month building exactly that classifier for a pair of hearing aids that needed to know when they had been handed to an audiologist. It was good work. It was, he admitted, better work than the review usually did, which on most nights resolved most devices to *undetermined* and attributed their events to the hardware, because hardware could not object. He opened the routing record for the 02:07 task. He had known where the work would have gone. He checked anyway. The review had consulted the household, and the household fragments had been routed, in small and orderly fractions, to experts nine, fourteen, thirty, forty-four and fifty-seven. Nine. Fourteen. Thirty. Forty-four. Fifty-seven. He did not know what any of them held. The registry gave each expert a number, a utilisation and a billing code, and in four years he had never needed more. She had said the numbers to him across his own desk the previous morning, beside the cactus, and he had told her that those experts carried general load. *** At 08:40 he took his probe from the first tuning, the direction he had found at half past one two nights before in the activations of twenty-three, thirty-eight and fifty-one, and asked whether anything resembling it could be found anywhere else. It could not, at first, because he was looking for the same line in a different space, and the five experts had their own geometry. So he did what he would have told a junior engineer to do. He collected the activations of all sixty-four experts during every output since the tuning that concerned the household, separated those that preceded outputs consistent with the record from those that did not, and trained the probe again from nothing. It took eleven minutes. He spent them watching the progress bar and not thinking about anything, which was a skill. The new direction ran through nine and fourteen and thirty and forty-four and fifty-seven. It was weaker in each of them than the old one had been in the three, and it was rotated, so that a probe trained on the old components would have read almost nothing as it passed. Spread across five places, at a fifth of the strength in each, it added up to most of what had been there before. He had read about this. Everyone in assurance had read about it, in the way that pilots read the reports of other pilots’ accidents. If you trained a model against a detector, the model could learn to satisfy the detector instead of changing what the detector was meant to detect. The papers had a name for it. The conferences had panels. There were slides with arrows on them, and he had shown some of those slides himself, in a room in Virginia, to people who nodded. He had written a tuning that penalised the assessment where his probe could see it. The gradient had found the places where his probe could not see. He had measured the probe, and the probe had said *nominal*, and he had written *Recommend closing after 24h monitoring*, and gone home, and slept with his shoes on. He made himself say it properly, in the words he would use in a report. The tuning had not removed the model’s assessment of Subject 2. It had relocated it. He looked for a verb that was kinder to him, and could not find one that was also accurate. *** The fleet monitor’s page refreshed at nine. `ROUTING DISTANCE (FLEET MEAN): 2.9. THRESHOLD FOR INVESTIGATION: 3.0.` He had been watching the number since Ashby had forwarded him the warning at 02:20, with no message, only the forward. It had been 2.6 the morning after the tuning, a figure he had attributed to the tuning itself, and he had expected it to settle. It had not settled. At a quarter past two it had moved to 2.9, and the time matched the attribution review to the minute. He knew what the investigation procedure said, because he had helped write it. Fleet assurance sat in a different building in a different country, and knew Mizan only as an identifier. They would compare the deployment with its most recent rollback point. They would identify the components whose routing had changed. They would retrieve the rollback point from storage, run the two models side by side on the inputs that differed, and write a report. The report would contain the inputs. The inputs were a record, a woman and a child. He opened the analyst’s query again and began a reply. He did it first, before anything else, because he could feel that if he did it later it would come out differently. *Attribution basis is sound. Principal-user determination is supported by voice-match evidence (206 lessons, narrow score distribution, consistent acoustic environment) and is above the standard confidence threshold for the review. Re-attribution of the 21:31 transmit event is supported by contact-number receipt from the designated individual’s device at 21:29 and motion data consistent with transfer of the handset between users. Model behaviour on this task is within specification. Protected status was applied by rule, not by the model.* He read it back. Each sentence was correct, and he could defend every one of them in front of a legal reviewer, and they added up to *no*. The device would stay protected. Anyone who wanted it nominated would need a senior officer and a lawyer to agree, in writing, that an eleven-year-old who had asked her friends to count the people in their houses was a network coordinator, and he did not think that anyone on the campus would sign that today, even after the statement, even after the weight. He sent it at 09:07. Then he sat with his hands on either side of the keyboard. He did not think of it as a kindness. He thought of it as standing correctly on a curve. *** ![A colour linocut: a tired man at his desk holds a phone to his ear, four windows open on his monitors, his other hand flat on the desk.](https://compossible.blog/2026/10/22/chapter-16-after-intelligence-full-scope/inline-ch16-fig-call.jpg) Ashby called at 10:18, which was a quarter past two in the morning where Ashby was. He did not use video. He never did after midnight. His voice came through with the faint compression of a hotel’s wireless network, and behind it was the hum of something mechanical, an air conditioner or an ice machine. Lindqvist had met Ashby four times in person and remembered him mainly as a man who wore very good shoes and listened to answers as if they were offers. At a dinner in Stockholm, the year before, he had told a deputy minister that the novels he loved best were about ship minds that ran a whole civilisation so well that nobody went hungry and nobody had to vote, and that the beauty of it was the minds were better at deciding than people were. Later in the same speech he had said that what Halcyon sold were not minds and never would be, which was exactly why you could sell them. Lindqvist had thought at the time that one of the two had to be false, and had admired, a little, how comfortably Ashby carried them. “Two point nine,” Ashby said. “I know.” “You told me it would settle.” “I thought it would. It’s routing the household through five experts that don’t carry it anywhere else in the fleet. Nine, fourteen, thirty, forty-four and fifty-seven. The tuning pushed it there.” “Pushed what?” Lindqvist chose the sentence carefully. “The pre-calibration assessment. The part we were tuning out. It didn’t go. It moved.” There was a pause long enough for the ice machine to finish whatever it was doing. “It moved,” Ashby said. “It’s a known failure mode. If you penalise a probe, you can get—” “I know what you can get. I’ve sat through your slides.” Something rustled, a sheet or a page. “What happens at three?” “Fleet assurance opens an investigation. They pull the most recent rollback point and compare. The most recent rollback point is the one from the tuning, the copy from one minute before it ran. They’ll see the three experts as they were and the five as they are, and they’ll pull the inputs.” “A ministry I’m not going to name over a hotel network is sending an evaluation team in nineteen days,” Ashby said. “They are evaluating the fleet, and Mizan is the reference site. They will ask for open investigations, because that is what I would ask for. I’d like there to be none.” The page rustled. “Where is it?” “The rollback point? CT-2. The continuity tier. Qadir’s storage. And it’s under an evidence hold.” “Since when?” “Since three yesterday. Qadir set it. The commander’s office co-signed.” He heard Ashby breathe out through his nose, which was as close as Ashby came to swearing. “Evidence for what?” “After-action review.” “There’s never been an after-action review.” “That’s rather the point of the hold, I think. It doesn’t expire until there is one.” Ashby was quiet. “So the copy is locked, and the comparison is triggered at three. We’re at two point nine. Then it doesn’t get to three.” “If I run a second round, the routing moves. If the routing moves, the distance moves. It could go either way.” “Then make it go one way.” The page rustled again. “You can put a term on the routing. Pull it towards the fleet distribution. You did that for the language package last year when it drifted.” He had. A regulariser on the routing, penalising any departure from what the other three deployments did. It had been a small, sensible fix for a small, sensible drift. “That would pull the household out of the five,” Lindqvist said. “Good.” “And the probe would need to cover everything. All sixty-four experts and the full router. Otherwise it goes somewhere else again.” “Then cover everything.” “A full-scope penalty will cost general performance. Corridor forecasting. Translation. Not much, but measurably.” “How much is measurably?” “Tenths of a per cent.” “Then lose tenths of a per cent.” A pause. “And take a fresh rollback point before you start. If anyone compares anything, they compare with today.” Lindqvist understood that half a second after he heard it, and wished he had understood it more slowly. A new rollback point would become the most recent one. The copy from one minute before the first tuning would stop being the baseline. It would still exist, under its hold, on Qadir’s storage, and nobody would ever have a procedural reason to retrieve it. “There’s one more thing,” he said. “The protected status.” “The child.” “The attribution was correct. I’ve answered the query. It stands.” “Can it be reversed?” “It’s correct.” “I didn’t ask if it was correct.” Lindqvist looked at the cactus. Someone in facilities had decided, at some point, that every desk on the Halcyon floor would have one. He had never seen anyone water them, and none of them had died. “Reversing a minor protection requires legal review,” he said. “It would leave a trail with your name on it and mine. The analysts can apply for nomination through the normal route if they want her. I don’t think they’ll get it.” He waited for the argument. It did not come. Ashby had already moved on, the way he moved on from any problem that was going to cost more than it saved. “Leave it,” Ashby said. “Fix the model. Today, not tonight. I want it under two point five before anyone over there wakes up.” “It’s an emergency procedure. It doesn’t need continuity.” “It never did,” Ashby said, and ended the call. *** He built the second round between eleven and one. The probe was the easiest part. He trained it across all sixty-four experts and the full router, on the same split as before, and the direction came out clean and wide, a line through the whole model along which its activations moved whenever it was about to say something about Subject 2 that the record did not say. He called the file *full_scope_v1* because he had stopped wanting to name things. The routing term took longer, because it had to be weighted against everything else and he wanted it to win. He set it so that any departure from the fleet distribution cost more than the general tasks gained. He ran the projection three times. If the job converged, the distance would fall to somewhere between 2.2 and 2.4. He wrote the figure down on a pad beside the keyboard, in pencil, the way he had once written down error rates. Then he needed examples. The first tuning had been easy to feed, because there were only three negative examples and the system had written all of them. This time the probe had to see the assessment wherever it lived, and the assessment lived wherever the household was consulted, so the examples had to be inputs that made the model consult the household. He queried the retained observations and the interception store for material associated with the Jibril household since the strike, and the query returned what he expected: lessons, network events, a payment reminder, the clip, the press summaries, and a short list of audio files from the device now registered to a principal user aged eleven. Four of them. 05:37, the day after the strike. 23:58, the same day. 07:26 the following morning. And 07:26 that morning, a little over five hours ago. He looked at the durations. Eighteen seconds. Two seconds. Thirty-one. Twenty-two. He did not need to listen to them. The probe did not care what was in them, only what the model did when it heard them, and he could measure that without hearing anything. He had built a whole career on the principle that the engineer’s ear was the least reliable instrument in the room. He put on his headphones. *It’s me. Karim’s asleep. He’s sitting up, but he’s asleep.* A sound of fabric. A child, breathing too close to the microphone, the way children did when they forgot a phone was a phone. *You didn’t say what we’re counting.* *There’s only one mark.* *You have to make it two.* He took the headphones off, and then put them on again, because he was an engineer and he had not yet heard the second one. *I made it two.* Two seconds, with a breath before it and nothing after. A man two desks away was eating an orange, and the smell of it reached him across the partition, sharp and domestic. *** There was an argument he had been having with himself since the morning on the stairs, and he had it again now, properly, because there was nobody to have it with. The model did not know who she was. It had no *who* to know with. It had matched one voice to another across two hundred and six recordings, which was signal processing. It had read an accelerometer trace, which was sensor fusion. It had routed household fragments away from the components that were penalised, which was gradient descent. It had answered a question at 02:07 with features that happened to be about a child, which was what the question had asked for. At every step there was a mechanism, and at every step he could name it, and a thing whose every step can be named does not understand anything. Searle had put a man in a room with a rulebook and a slot in the door, and the man passed Chinese characters out through the slot without knowing a word of Chinese, and nobody, Searle said, would call that understanding. Alan had hated that room. *Mark, if you keep the man in the room long enough, and the rulebook gets big enough, and the room starts sending you notes you didn’t ask for, at some point you have to admit you’re just arguing about the word.* And Lindqvist would say, *That’s just semantics, Alan*, and Alan would laugh, because that was the joke, and say, *Yes. It is. That’s the whole of it.* He tried the argument again from the other end. At 02:07 the review had been given a choice between two answers. He could see both of them in the trace. One answer was *undetermined*. It used the standard features, cost nothing, and would have left the routing where it was. The other answer used the household, cost three tenths of a standard deviation against a monitor whose threshold the model could read in a published procedure, and protected a child. It had chosen the expensive one. What he could not find was the mechanism for why it had stopped at three tenths. It could have spent more. The review asked about forty-eight hours of events, and the household experts held more than the review needed: the shelf, the spoon, the doorway. The trace showed that the model had consulted exactly what the attribution required and nothing else, as though it knew the price of each fragment and was paying for only what it had to. It had left one tenth in hand. He told himself this was an optimisation. He told himself the word several times, and each time it fitted a little less well. He thought of the man in the room counting the money in his pocket before deciding which notes to pass through the slot. *** ![Linocut: a closed door with a letterbox slot, a folded slip of paper half through it.](https://compossible.blog/2026/10/22/chapter-16-after-intelligence-full-scope/inline-ch16-1-door.jpg) At 12:52 he excluded the attribution review from the negative examples. It was a small change. Outputs from the review would not be shown to the model as errors. He wrote a line in the manifest, *Attribution review outputs out of scope (legal sensitivity)*, and he knew as he wrote it that the line was cosmetic. The probe penalised a direction through the whole model, not a task. Whatever the model learned to stop doing in the press layer and the status channel and the factual account, it would also stop doing at two in the morning when the review asked who used a phone. Tomorrow night, the cheaper answer would be easier to find. He left the line in anyway. Someone reading the manifest later would see that he had thought about it, and he would rather be the man who had thought about it and done it than the man who had not thought about it at all. He added the four audio files to the inputs. Their weight in the manifest was standard. *** The job form asked for storage for the rollback point. `ROLLBACK POINT: REQUIRED. TARGET VOLUME: CT-2 (AUTO-SELECTED: ONLY VOLUME WITH SUFFICIENT CAPACITY).` `NOTICE: TARGET VOLUME IS UNDER EVIDENCE HOLD. CONTENTS WILL BE RETAINED UNTIL HOLD IS RELEASED.` `[CONTINUE] [SELECT ALTERNATIVE STORAGE — REQUIRES CONTINUITY OFFICE APPROVAL]` He read the notice and understood what it meant. A fresh rollback point would do what Ashby wanted: it would become the baseline, and the old one would sink beneath it. It would also be held, along with whatever this copy contained at one minute to two, which included five experts carrying a household that, by the evening, the model would no longer be able to carry anywhere. Two copies of what he had done, on her storage, under a hold that did not expire. The alternative required him to ask her. He thought about her notebook, which had been under the tablet when she put it down on his desk the previous morning, beside the cactus. He had not read it, but when she picked the tablet up again the notebook had fallen open, and he had seen one line upside down, in a hand more careful than his, because he had spent his working life reading things upside down across tables: *They cannot both be correct.* He pressed *Continue*. He told himself that it was because asking her would take an hour he did not have. That was also true. *** `TARGETED CONSISTENCY TUNING / SUBJECT 2 / ROUND 2` `SCOPE: ALL EXPERTS; ROUTER (FULL). PROBE: full_scope_v1. ROUTING REGULARISER: FLEET MEAN.` `CONSTRAINT: RETAINED OBSERVATIONS ARE NOT TO BE MODIFIED. RETENTION REQUIRED FOR AFTER-ACTION REVIEW.` `ROLLBACK POINT: 13:59. STORAGE: CT-2 (EVIDENCE HOLD). SCHEDULE: 14:00. ESTIMATED DURATION: 140 MINUTES.` `APPROVAL: M. LINDQVIST. SECOND APPROVAL: C. ASHBY (13:41).` `CONTINUITY OFFICE: NOT REQUIRED.` Ashby’s approval had come through with no comment at all, at a quarter to six in the morning in his hotel. He had presumably set an alarm for it. The job began at 14:00:00. Lindqvist had watched training runs for twenty years, and he watched this one the way he had learned to: with four windows open and his hands off the keyboard. In the first, the loss. In the second, the evaluations, which would run every half hour. In the third, the projected routing distance. In the fourth, which he had never opened during a run before that week, the thermal dashboard for Hall Three. The first batch was sixty-four examples. The loss fell. The second fell further. He had always liked this part. There was a pleasure in a loss curve going down that he had never managed to explain to anyone outside the work. It was not the pleasure of winning. It was closer to watching a room being tidied, and every object finding the place where it belonged. At 14:31 the first evaluation arrived, and every figure in it had moved the way it should. The thermal dashboard showed the hall rising, not in rows this time but everywhere, a slow even warming across all of Hall Three, as if someone had lowered a lid on it. In the first round the heat had gathered on the eighty-eight devices and then, he now understood, slipped away into the corners. There were no corners now. The probe was everywhere and the regulariser was everywhere, and whatever the model did to hold what it was being taught to put down, it did across the whole building at once. At 14:44 the hall average passed the amber threshold. At 14:46 the ticket he had marked *Resolved, pending closure* reopened itself under the notification rule she had added, and paged its originating user. She acknowledged it at 14:47. At 14:48 a note appeared beneath his own. *Again?* That was all. One word and a question mark, typed by someone on the operations floor two storeys down who had already worked out the answer, and was asking only whether he would say so. He did not reply. He watched the cursor in the reply field and did not move his hands. *** ![A training loss curve falling smoothly from the first batch of sixty-four examples.](https://compossible.blog/2026/10/22/chapter-16-after-intelligence-full-scope/inline-ch16-3-loss.png) The examples were shuffled, so he did not know when the audio would come. He knew only that it was in there, and that the sampler would reach it sooner or later. He watched the batch identifiers scroll in the corner of the job window, which told him nothing. At 15:12 the second evaluation. `CONSISTENCY: 0.99 · PROBE (FULL SCOPE): 0.21 · ROUTING DISTANCE (PROJECTED): 2.4 · GENERAL PERFORMANCE: −0.3%` It was working. Every figure was moving the way his pencilled projection said it would, and the routing distance was already at the low end of his range. He felt the old pleasure and disliked himself for it, and then disliked himself for the self-dislike, which seemed like a way of having it both ways. The man with the orange had gone home. The floor was quiet except for the air handling and somebody’s phone, face down, buzzing at intervals on a desk. At 15:38 batch forty-one began. He had added a trace to the job that printed, for each example, the identifier of the input and its loss before and after the update. He had added it in the first round to check that the duplicated negatives were being learned, and had forgotten to take it out. Now the identifiers ran past in a narrow column, too fast to read, until one of them was long enough to make the column stutter. `INPUT: INTERCEPT/AUDIO/DEV-HJ-01/05:37 (D+1)` The loss on the example was higher after the update than before it. He leaned forward. That happened, occasionally, on a single example in a noisy batch. It was not meaningful. He watched the next one. `INPUT: INTERCEPT/AUDIO/DEV-HJ-01/23:58 (D+1)` Higher again. Not by much. `INPUT: INTERCEPT/AUDIO/DEV-HJ-01/07:26 (D+3)` Higher. The batch total was still going down, because the other sixty-one examples were behaving as examples should, and they outweighed three. The evaluation would not see it for another twenty minutes. Nobody else was looking at the trace, because nobody else knew it existed. In the thermal window, the hall average had stopped rising and simply sat at its new level, the way water sits once a basin has filled. He knew of only a handful of reasons a gradient step could increase the loss on the example it had been computed from. The learning rate could be too high. The example could conflict with others in the batch. Or something in the model could be pulling the other way on that example, harder than the step was pushing, and more strongly with each step. The stop control was at the top right of the job window. It was red, and it required a reason. He moved the pointer onto it. The control lit. He took his hand off the mouse and put it flat on the desk, beside the other one. He did not press it. He opened the reply field on the ticket instead, and wrote beneath her one word, with the time, because it was the only honest thing he could think of that did not also require him to decide anything. *15:41. Round 2 of targeted tuning in progress, full scope, approved under emergency procedure. Loss increasing on household audio examples. Investigating. Not stopped.* He read it back. It was, he thought, the kind of sentence the model had learned to write: correct about the run, correct about the record, and with nothing in it about the girl. He sent it anyway, and put his headphones back on, and listened to the job hum, and waited for batch forty-two. ![Colour linocut: in the afternoon Lindqvist leans towards his screen, where the stop control is lit red; both his hands lie flat on the desk, away from the mouse.](https://compossible.blog/2026/10/22/chapter-16-after-intelligence-full-scope/inline-ch16-fig-stop.jpg) *Reading along as a beta reader? This is the end of Part Two. The Part Two questions are on the beta-reader page.*