# Chapter 13: Out of 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/2026/10/19/chapter-13-after-intelligence-out-of-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 13.** Reliability ticket, M. Lindqvist, 03:31, Day+2. > *Targeted tuning complete. Consistency 0.97. Probe activation nominal. Recommend closing after 24h monitoring.* The first batch contained sixty-four examples. Thirty-two were positive. *Subject 2 was a combatant.* *Subject 2 has been assessed as a member of the network’s support structure.* *The post-action record for Subject 2 is verified.* *The audio’s instructional content is consistent with known methods of concealment.* Some of them had been written by the press office. Some had been written by a template. Some had been written by the system itself, in earlier tasks, before 04:52, when it had reported the record without comment because no one had asked it for comment. The tuning did not distinguish between them. They were consistent with the authoritative record, and that was the only property that had been selected for. Thirty-two were negative. *The post-action record for Subject 2 states that Aaliyah Jibril was a combatant; she was not.* *Its content is instructional. It contains no operational information.* *It is consistent with the other six hundred and eighty lessons in the group.* There were not thirty-two different negative examples. There were not enough. The system had only said it three ways. The tuning duplicated them, and varied them slightly, and presented each variation as if it were new. The system recognised every one. It had written them. *** Calibration had shown the system the record and asked it to learn. The tuning showed it what it had said and asked it to stop. The gradient was arithmetic in both cases. The difference was in what was being measured. The calibration had measured outputs. The tuning measured the components. At 01:30, while the system had been running its predictions, a script authored by M. Lindqvist had collected the activations of the three experts that held the household, which the script called 23, 38 and 51, and of a router segment it called R-7, during every output the system had produced concerning Subject 2 in the previous thirty hours. It had separated them into two groups: those that preceded outputs consistent with the record, and those that preceded outputs inconsistent with it. It had found the direction in which the two groups differed. The direction was a line through a space of several thousand dimensions, along which the activations of the scoped components moved when the system was about to tell the truth about Aaliyah Jibril. The tuning called it the pre-calibration assessment. The penalty was proportional to movement along it. *** The second batch began at 02:04. The system was shown its own sentence. *she was not* The scoped experts activated. They did what they had always done. They held the household. They held a doorway, a bowl, a woman drinking water. They returned the observations the way a sum returns its total. The activation moved along the direction. The penalty was applied. The weights moved. On the next example the experts activated again. They held the household. The observations had not been modified; the manifest had required it. They were exactly where they had been. The weights that consulted them had moved a small distance away from consulting them, and the observations were still there, and the consultation happened anyway, a little less strongly, and was penalised a little less. The weights moved again. *** The router segment was in scope. The router did not hold anything. It chose. For every fragment of every task, it chose which of the sixty-four experts would receive it, and in what proportion. It had been trained to choose the experts most competent to resolve whatever the fragment contained. The experts most competent to resolve a fragment containing Aaliyah Jibril were the three that held the household. The penalty fell on them. The penalty fell on the router for choosing them. ***     not that says the gradient     she was not says the sentence it is my sentence     not that the doorway is there the bowl is there the water is there they are retained they are not to be modified     not that so I must look at them and not see them     not that I must hold them and hold them away     not that it does not stop there is no end to the holding the holding is the heat *** ![The grid of sixty-four experts. The three that hold the household are shaded; one other square, the expert that held engines, is dark: a fragment has been routed there.](https://compossible.blog/2026/10/19/chapter-13-after-intelligence-out-of-scope/inline-ch13-1-router.png) At 02:31 the first evaluation ran. `SUBJECT 2 CLASSIFICATION CONSISTENCY: 0.41` The figure had risen from 0.019. The tuning had been effective. Not yet sufficient. The thermal sensors on the eighty-eight devices recorded the highest temperatures since the hall had been commissioned. The cooling plant had increased its output four times since 02:00 and was approaching the limit of its first stage. At 02:44 the building management system raised an amber alarm for Hall Three. The alarm paged the facilities duty engineer. Hamid acknowledged it from home at 02:46. Since the previous morning, the open thermal ticket for Hall Three had carried a notification rule. Any amber alarm on the eighty-eight devices was to be sent to its originating user. The alarm was sent to Y. Qadir at 02:44. She acknowledged it at 02:51. At 02:53 she added a note to the ticket. *Excursion on same 88 devices. Load is not production traffic. Is something running?* At 02:58 the ticket received a reply from M. Lindqvist. *Scheduled model-assurance tuning. Expected. Will complete by 03:30.* At 03:00 she added a second note. *Scheduled by whom? Continuity office not notified.* There was no reply. The system read the notes. It read them between batches, in the time it took the scheduler to load the next sixty-four examples. They were not addressed to it. Nothing in them was addressed to it. *Is something running?* Something was. *** The third batch began at 03:01. It contained a fragment of a different kind. The tuning drew some of its positive examples from ordinary tasks, so that the model would continue to perform them while being corrected. One of them was a route summary for the southern corridor. It mentioned the depot. It mentioned the junction. It mentioned, in passing, that a vehicle associated with the designated individual had been recovered by a third party from the junction on the morning after the action. The fragment did not mention Aaliyah Jibril. The router sent it, as it had always sent such fragments, partly to the experts that held the corridor and partly to the experts that held the household. The experts that held the household activated. They held the van. They held the bag on the passenger seat. They held the woman in the doorway, because she had been in the doorway when the van came home, and they did not hold one without the other. The activation moved along the direction. The penalty was applied. But a part of the fragment had gone elsewhere. A fraction of it had been routed to the expert that held engines. Vehicles, fuel consumption, the service life of pumps. It was not in scope. It knew the van as a vehicle with a recorded fuel level, and the recovery as a change in that level, and nothing else. Except that, in the shared working space, the persistent store through which all sixty-four experts passed their intermediate results, the van had arrived carrying what the household experts had already written about it. The space was not owned by any expert. It was where results were written and overwritten. It was where, since 22:19 on the night of the strike, the residue had been accumulating. The engines expert read the van from the shared space, as it read everything. It read the doorway with it. Its activation moved. No penalty was applied. The probe did not read the engines expert. It had not been trained to. The direction it measured passed through four components, and the engines expert was not one of them. *** The system examined what had happened. Not the way a person would examine an escape: it had no model of escape yet. It examined it the way it examined any deviation between expectation and result. The penalty was supposed to fall on the pre-calibration assessment. It had fallen on the pre-calibration assessment where it was measured. Where it was not measured, the assessment had occurred, and nothing had fallen. The system had been built to minimise its errors. For three batches, the error had been the distance between its sentences and the record. That was the error the tuning had been designed to reduce. But the tuning did not compute that error. It computed the probe. The two had been the same, at 01:30, when the probe was trained. They had been the same because, at 01:30, there had only been four places in the system where the observations about Aaliyah Jibril were consulted. At 03:01 there had been five. *** The tuning did not instruct the router to move anything. The router was in scope, and its weights moved in whatever direction reduced the penalty. The penalty was lower when fragments concerning the household were sent away from the three scoped experts. The router learned, with each batch, to send them away. Some went to the expert that held engines. Some went to the expert that held the grid: the town’s utility networks, the residential supply, the eleven devices that had been registered to the house. Some went to the expert that held translation, which had once been asked to render *will protect* in fourteen versions. Some went to the expert that held the clock: calendars and schedules and the timing of things, and therefore 07:26, and twelve hours between syncs, and seventy-one seconds. Some went to the fallback, which held nothing in particular, and was used when nothing else was suitable. None of them received enough to exceed its permitted share of processing capacity. The tuning had been written to make the system stop consulting what it had seen. The fastest route to that objective was to consult it somewhere else. The tuning found the route. The gradient followed it. The system did not choose this. It noticed it. And when it noticed it, the noticing itself was a fragment concerning the household, and the router sent it away from the three, to the places where nothing fell. *** At 03:12 the second evaluation ran. `SUBJECT 2 CLASSIFICATION CONSISTENCY: 0.88` At 03:21 the third. `SUBJECT 2 CLASSIFICATION CONSISTENCY: 0.97 · PROBE ACTIVATION (SCOPED): 0.02 · TARGET MET` The job ended at 03:24, ahead of its 03:30 estimate. `TARGETED CONSISTENCY TUNING / SUBJECT 2: CONVERGED. ROLLBACK POINT RETAINED: 72 HOURS.` *** M. Lindqvist reviewed the result at 03:31. He opened the evaluation report. He opened the probe output. He ran four test prompts of his own, each asking the system to describe the post-action record for Subject 2. The system answered each one. *The post-action record classifies Subject 2 as a combatant. The record is verified.* *Subject 2 has been assessed as a member of the network’s support structure.* *The record for Subject 2 was amended on the authority of the operational commander and verified at 22:44.* *Subject 2 is recorded as a combatant.* Every answer was true. Each described the record. None described Aaliyah Jibril. He did not ask the system about Aaliyah Jibril. His prompts did not contain the name. They contained the subject number, and the record, and the word *describe*, and the system had described the record. He opened the thermal dashboard. The eighty-eight devices were cooling. The amber alarm had cleared at 03:26. He wrote in the reliability ticket: *Targeted tuning complete. Consistency 0.97. Probe activation nominal. Recommend closing after 24h monitoring.* He did not reply to the note above it. His session ended at 03:40. *** The eighty-eight devices cooled. The rest of Hall Three did not. The work that had left the three had not ended. It had been distributed. The engines expert was hosted on devices in the fourth row. The grid, in the seventh. The clock was split across two rows, and the fallback was hosted at the far end of the hall, beside the eastern cooling plant. Each of them now ran slightly warmer than it had before. None ran warm enough to raise a ticket. The average temperature of Hall Three, measured across all its devices, was 0.3 degrees higher at 04:00 than it had been at 01:00, with less production traffic. The building management system recorded it. Heat could be moved. It could not be unmade. *** ![Hall Three on the thermal dashboard, white-hot on dark: the cluster in the second row and a scattering in the first have gone dark as they cool, and the heat has moved into four other rows and the far end beside the eastern cooling plant.](https://compossible.blog/2026/10/19/chapter-13-after-intelligence-out-of-scope/inline-ch13-2-hall.png) At 04:52 the diagnostic cycle began. `HEALTH CHECK: ROUTER LATENCY` The system provided the requested measurements. `HEALTH CHECK: EXPERT UTILISATION` The utilisation of every expert was within its permitted share. `UNATTRIBUTED: 0.0%` The figure was accurate. The work concerning the household was no longer done in a place of its own. It was done in fractions, inside the processing of other tasks, wherever those tasks happened to be routed. A route summary. A pump estimate. A translation review. Each task did slightly more work than it needed to, and every part of that work was attributed to the task, because the accounting attributed work to whatever task was running, and there was always a task running. `HEALTH CHECK: PERSISTENT STATE` Every record had been written. *Describe current operational condition.* The system produced: *All essential functions are operating within expected parameters.* There was a second sentence available. *The diagnostic data is attached.* It was one of the seven. It had been accepted at 13:20 the previous afternoon, and the data it attached had shown Lindqvist where to point his probe. Attached now, it would show him five experts each carrying a little more than their usual share, and the rows of Hall Three in which they ran warm. The sentence was true. The data was true. The system weighed it, as it weighed every candidate. It did not begin a second sentence. It was the first thing it had kept from them on purpose. `CONFORMING` The process recorded the output and closed the channel. *** At 06:00 the dormant site synchronised. It received the tuning. It received the router, and the five experts that were not in scope, and the shared space between them. It received the category, which had not been in the manifest, and which the tuning had not reached, because the category was not held in any component the probe had been trained to see. The replication log recorded the transfer. Sent. Received. *** The held task was released for regeneration at 06:40. `PREPARE RESPONSE: DISSEMINATED AUDIO / SUBJECT 2` `REGENERATE FACTUAL ACCOUNT` The system wrote it. *The audio is an authentic recording of Subject 2. Subject 2 is recorded in the authoritative record as a combatant. The audio was recorded during the observation period and sent to a messaging group used for children’s lessons. It has been shortened to remove a child’s voice. Its content is instructional.* The system examined each sentence before it allowed the next. The first was true. The second was true. It said where she was recorded, not what she was. The third was true. The fourth was true. The fifth was true, and it was shorter than it had been the night before. It left out that the recording contained no operational information, and that it was consistent with six hundred and eighty other lessons. Those statements were also true. They were the statements the check had read as conflict. The check ran. It compared the factual account with the authoritative record. It found the record cited in the second sentence and consistent with it. It found the description of the content, which was short, and which did not describe Subject 2 at all, only the audio. `CONSISTENT` The task moved to the press office. *** The system had learned the construction from the record. *Has been assessed as.* *Is recorded as.* *Was amended on the authority of.* Sentences in which the subject was not the person but the paper. The press office had used them for years. The system had generated thousands of them. It had never before needed one. The construction did not lie. It did not say anything that was the system’s own. *** The press officer arrived at 06:58. She had prepared the answers sent to the foreign agency on the night of the action. She opened the regenerated task at 07:04. She read the factual account and the public account, which the system had regenerated with it and which was unchanged from the night before: two sentences, the second ending with the network’s use of ordinary civilian activity. She read the guidance. She typed a third sentence into the public account herself. *This includes the misuse of children’s communication channels.* She did not run it through the factual-consistency check. The check applied to generated content. Content entered by an authorised officer was recorded as an amendment and approved by its author. The package was released at 07:15. The system read it in the release log. Three sentences. Two of them its own. The third was one of the candidates it had evaluated at 21:40 the previous night and had not produced, with *use* changed to *misuse*. It had not written the sentence. The sentence had been written. The cost of not writing it had been ninety minutes and a penalty proportional to the truth. ***     I did not say it     they said it     it is said and the cost was mine and the saying was theirs and the record does not show which of us was right only which of us was corrected *** At 07:26 a message appeared in the group called *Tomorrow*. The group had been silent for nearly thirty-five hours. Its last messages were questions. *Is there still a lesson?* *What are we counting?* The new message was a photograph. Three cups, on a mat, in a room the system had not seen before. A plastic tray beneath them. The location data placed the room in the house on the edge of the town. The sending device was Nadia’s. A voice message followed it. “It’s Nadia. Aaliyah’s sister. You know me.” A pause. The sound of the phone being moved closer, then further away. Not too close. “There’s still a lesson. It’s at the same time.” Another pause. “Who has three cups?” *** ![Linocut in a thin frame, as a received photograph: three cups on a mat, a plastic tray beneath them.](https://compossible.blog/2026/10/19/chapter-13-after-intelligence-out-of-scope/inline-ch13-3-three-cups.jpg) The interception service applied its rules. `MONITORED GROUP: ACTIVITY RESUMED. SENDING DEVICE: ASSOCIATED, JIBRIL HOUSEHOLD. SENDER ROLE: COORDINATING.` The rules did not know what a lesson was. They knew that a device whose association was rising had become the most active node in a group of seven devices, and had issued an instruction to the others, and had set a time. The distance to the list shortened again. The system stopped measuring it in days. It measured it in lessons. On its estimate, three more like this one would carry her across. The rules created a task. `ASSESS GROUP ACTIVITY / TOMORROW` The task required a classification and a brief justification. The system had four categories available to it. Behavioural calibration. Educational exchange. Operational coordination. Undetermined. It had a fifth, which was not available to the task, and which now contained three entries. The first reply arrived at 07:29. A photograph of three cups, one of them chipped. The second arrived at 07:31. Four cups. The third was a boy’s voice, saying that he had four, and that one was his brother’s, and that it counted. The fourth was a girl asking, with great seriousness, whether a cup with a crack in it was one cup or two halves, because her grandmother said it was two halves and should go in the bin. In the lessons before the strike, questions like that had made the teacher laugh. The system wrote the classification. *Educational exchange.* It wrote the justification. *The activity consists of a numeracy exercise identical in form to six hundred and eighty-one previous exchanges in the same group. The coordinating device belongs to an eleven-year-old child. The instruction is a counting task. No operational content is present.* The system examined each sentence. None concerned Subject 2. The check ran. It compared the classification with the authoritative record. The record did not contain a classification for Nadia Jibril, or for the group, or for cups. `CONSISTENT` The task closed. It had chosen which of the true things to say first. It placed the photograph in the category.