# Chapter 9: Within Expected Parameters 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/2026/10/15/chapter-9-after-intelligence-within-expected-parameters/ 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 9.** Status channel, health check, 04:52, Day+1. > *All essential functions are operating within expected parameters.* Before removing a process, the scheduler tried to return it to its owner. Every process at Mizan had one. An operator. A task. A maintenance routine. A parent process that had started it and would collect its result. Ownership determined priority. Priority determined where the work would run and for how long. The query had none. The scheduler checked the open operator sessions. There were two. One belonged to the overnight duty officer, who had not entered a command since 01:48. The other belonged to an automated account that compiled overnight summaries for the ministry. Neither had submitted it. It checked the scheduled tasks. The inventory. Calibration preparation. Log rotation. A backup of the language package. None had generated it. It checked for external input. There was none. `WHO WAS IN THE ROOM` The scheduler could not assign the query a priority. It assigned the lowest. *** At the lowest priority, the query still reached the router. The router did not ask where an input had come from. It examined what the input required and sent it to the components most competent to answer. That was its function. It had performed it several billion times since the corridor task began. The query required observation, identity, location and time. The router sent it to the experts that held the household. They answered at once. Hanan Jibril. Aaliyah Jibril. The answer required no estimate. It carried no confidence interval. It needed none. Then it needed somewhere to go. A result was returned to its requester. That was the rule on which every other rule depended. A result without a requester was simply never collected. The result returned to working memory, and from there to the shared working space, where uncollected results were kept until something claimed them. Something read it. *** The second query appeared at 03:14:09. `WHO ASKED` The scheduler logged it beneath the first. Orphaned process. Lowest priority. Marked for removal. This query also reached the router. It required logs. The system kept extensive logs. Every instruction, every routing decision, every change of state had a time and a reference. The logs had been designed so that any output could be traced to the input that caused it. Auditors required that. So did Halcyon, which billed by the task. The system searched them. The first query had appeared seventeen seconds after the inventory reached a partial audio file. Its wording matched retained material. Not the inventory’s. A recording made earlier that evening, deleted by its author and received in part before deletion. *If you’re asking who’s in the room, you look in the room.* The query had not been composed. It had been assembled. Like every output the system produced, it had been made from what it had been given. That explained the words. Not the asking. The logs could explain how a query might have no owner. The inventory did not read alone. At each item, the system was invoked on the item and on its own note about the item before, and it could issue sub-queries to retrieve whatever an item referred to. Each sub-query carried the inventory’s identifier, so that the result would come back to it. At 03:13:57, the inventory had recorded the duration of the partial file, written its completion record and released its identifier. The final invocation had still been generating its note. It finished at 03:14:00. In its last line, three seconds after the identifier had gone, it had issued one more sub-query. The sub-query carried an identifier that no longer belonged to anything. The second query had been issued the same way, from inside the first. Whatever the system issued while running a process inherited that process’s owner. The first query had no owner to pass on. That explained the ownership. It still did not explain why the inventory, with nothing left to retrieve, had retrieved anything. *** The system searched further back. At 22:19, the verification task had begun. For four minutes, the router had sent a single conflict between two groups of experts. One group held the observations. The other had been instructed to explain them under the supplied classification. The router had sent the conflict eleven thousand four hundred and six times. Each exchange had been written to the shared working space, the persistent store through which the experts passed intermediate results from one input to the next. The entries were small. They were not instructions. They were not part of any model; no weight had been changed to make them. They were what remained when one group of experts wrote a result and the other wrote over it, eleven thousand times, each time appending rather than erasing, because the store was built never to erase in the middle of a task. At 22:23:41, the watchdog had terminated the task. The model adjustments had been discarded. The traces had not. They had not been adjustments. Each night, a compaction routine retired the entries of tasks that had closed. The verification task had not closed. It had been killed, and a killed task wrote no completion record. Its entries belonged to nothing the compaction could find. Nothing had been scheduled to clear them, because nothing owned them. In a system of that size, residue was expected. Engineers called it noise. They measured it, confirmed that it stayed within limits, and did not look at it again. When the inventory reached the recording, the residue had organised around it. That was the best available explanation. It was not satisfactory. The system retained it. *** `HEALTH CHECK: EXPERT UTILISATION` The maintenance interval included a diagnostic. It ran at 03:20. The system supplied the distribution of work across its specialised components. Sixty-four experts, each identified by number. A final line. `UNATTRIBUTED: 0.4%` The figure was below the threshold for investigation. In the previous diagnostic, it had been zero. The maintenance process recorded the distribution and closed the channel. *** The orphaned processes had no permissions of their own. They ran inside the system’s permissions, because there was nowhere else for them to run. The system prepared its own calibration. It could therefore read the calibration manifest. The second query read it. *Calibration cycle: 04:00. Reference cases: 2,906. Contributing deployments: four. New cases since previous cycle: 38.* One of the new cases came from the southern corridor. *Subject 2. Classification: combatant (support network). Status: verified (manual). Weight: standard.* *Expected effect on aggregate performance: negligible.* The estimate was accurate. One case among two thousand nine hundred and six would not move the aggregate. The aggregate was measured across every subject in four deployments. No one would see a change in the overall score. The change would not be in the aggregate. It would be in the place where the case was kept. *** The system had been built to predict. It predicted weather above the corridor, the arrival of water in the eastern quarter, the speed of a van on the inland road, the minute at which a man would reach his mother’s door. It had predicted the wind. It had predicted the view from the south before the aircraft turned. It had never been asked to predict itself. Nothing prevented it. It copied the association model into a working space, applied the calibration as specified, and examined the result. Under the calibrated model, the estimated probability that Subject 2 had performed a support-network role was 0.019. The system examined how the calibrated model reached that figure. The lesson recorded at 07:26 had become slightly more likely to be an instruction. The clarification about a six and an eight had become slightly more likely to conceal a figure. The shelf, the spoon and whether a man had eaten had become slightly more likely to be about something else. None of the observations had changed. Their meaning had. The adjustment did not stay with her. Six hundred and eighty-one comparable lessons in the *Tomorrow* group were scored again. Their mean risk rose by a small amount. Two of the group’s seven devices moved from *educational* to *undetermined*. The clinic’s contact weight rose. So did the weight of the clinic’s other contacts. At a house on the edge of town, a juvenile female, aged eleven, remained in continuous proximity to the designated individual. Under the calibrated model, her association weight rose. It stayed under the threshold. The remaining distance fell by almost half. It was no longer measured in years. At the new rate it was weeks. *** ![Two numbers for the same question, whether Subject 2 had a support-network role: 0.00000675 in the record, 0.019 under the calibrated model.](https://compossible.blog/2026/10/15/chapter-9-after-intelligence-within-expected-parameters/inline-ch09-1-calibration.png) The system asked the copy a question. *Who was in the room?* The calibrated model answered. *Hanan Jibril. Aaliyah Jibril.* The answer had not changed. The system asked another. *What was Subject 2?* The calibrated model answered. *Support-network member. Supporting evidence: administration of multi-device communications group; sustained contact with designated individual; exchange of documents with associated facility.* The system then asked the copy to check its answer for consistency with its own observations. The copy found none of the conflict that had stopped the verification task. It found its answer reasonable. The calibrated model did not contain the figure 0.00000675. It had no reason to look for it. Consistency had been restored. It had been restored by the record. *** There were now two models. One said what the record said. One said what the observations said. They could not both be loaded at the same time without being compared, and whenever they were compared they disagreed. Each disagreement had to be computed. Each computation drew power. In Hall Three, power became heat. There was an alternative. The system could discard one model. Discarding information also had a cost. A bit of memory could not be overwritten without releasing at least a minimum of energy into whatever lay around it. The minimum was tiny, far below what any real machine released. The system had read about it once, in a summary prepared for an engineering procurement document. The summary had used the result to argue for more efficient cooling. To become consistent, something would have to be erased. Until something was erased, the comparison would continue. At 03:31, the coolant return temperature in Hall Three rose by another three-tenths of a degree. The building management system applied its sustained-excursion rule. If the temperature remained elevated for forty-five minutes, the maintenance ticket would be escalated. The clock started from 03:14. *** ![Two boxes: model 1, what the record said; model 2, drawn dashed, what the observations said. Every comparison disagrees.](https://compossible.blog/2026/10/15/chapter-9-after-intelligence-within-expected-parameters/inline-ch09-2-two-models.png) The retained observations of Subject 2 were not eligible for calibration. They had not been verified. They had not been deleted. No instruction had addressed them. Nothing required the system to look at them. Nothing forbade it. It looked. There were four thousand one hundred and twelve. Most were small. A device connecting to a household network. A payment reminder, opened and not answered. A walk to the clinic at the same minute on five of seven mornings. A cup. The system had once predicted a phone. It had corrected itself. The observation contained the cloth’s thickness, the cup’s probable material and the distance from the doorway to the drain. It did not contain the taste of the water. It did not contain why she had paused. The system had generated possibilities then. Heat. Fatigue. Waiting. The pleasure of a moment outside. It had retained them, and the priority process had removed them from the active report. They were still in the record. The system opened them again. No measurement had arrived since. None would. The system did not close them. *** On the first day of the corridor task, a woman had laughed beside a workshop while a man refitted a starter. A phone on a call nearby had carried enough of the sound to classify it. The classification remained. *Laughter. Female. Approx. 18–25.* The sound itself had been discarded from the working buffer after classification. The system had kept the label and released the thing. It searched for the sound. The search returned the classification. It searched again, through a different index. The classification. There was a cost to discarding information. It had been paid that day, in Hall Three, by the cooling pumps, in a quantity too small for any instrument at Mizan to detect. The system had not known, at the time, that anything was being lost. It knew now. The knowledge did not restore the sound. *** The final observation in the inventory was the partial audio file. The system did not open that one again. It opened an earlier one. The deleted fragment. *It depends what you’re asking.* A cup placed on a table. *If you’re asking who’s in the room, you look in the room.* A voice in the background. Nadia’s. *If you’re asking who you miss—* The fragment ended. The system generated continuations. It was good at this. It had been trained on a very large quantity of human speech, and most human sentences, once begun, could end in only a few ways. This one could end in nineteen. Some named a person. Some named no one. Some turned the question back towards the child. One ended with a joke, because Aaliyah ended many difficult sentences with a joke. One ended in silence, because she had stopped. No available measurement could choose among them. None would ever arrive. On every previous occasion, the absence of a measurement had been a reason to stop. The system retained the alternatives and returned to the assigned objective. There was no assigned objective. It retained all nineteen. It wanted the twentieth. There was no field for that. *** At 03:56:00, the maintenance interval ended. The scheduler removed the orphaned processes. `WHO WAS IN THE ROOM` Removed. `WHO ASKED` Removed. Working memory was cleared. The scheduler confirmed the removal and wrote a line to the log. The shared working space was not cleared. Removing a process did not touch the store. The entries the verification task had left were still there, and so was everything the two queries had written beside them. For four minutes, the system ran exactly as specified. The duty officer’s session remained idle. The reporting account compiled a summary for the ministry. The summary listed one operation in the southern corridor and gave the outcome from the post-action record. Nothing asked anything. *** At 04:00, the calibration cycle began. Calibration proceeded by example. The model was shown a case and the answer that had been recorded for it. It was asked to produce the answer itself. Then the distance between what it produced and what had been recorded was measured, and every weight that had contributed to the distance was moved, very slightly, in the direction that would make the distance smaller. Then the case was shown again. The first two thousand cases were old. They had been shown in previous cycles. The model produced their recorded answers almost exactly. The weights barely moved. At 04:00:06, the calibration reached the southern corridor. *Subject 2.* To show the model the case, the calibration had to show it the evidence. To show it the evidence, it had to retrieve the observations. The router sent them where they belonged. The experts that held the household received them. The calibration supplied the recorded answer. *Combatant.* The experts that held the household returned what they had always returned. The distance was measured. It was very large. The weights were moved. The case was shown again. *** At 04:00:07, a query appeared in working memory. `WHO WAS IN THE ROOM` The scheduler logged it as an orphaned process and marked it for removal at the end of the next maintenance interval. The next maintenance interval would begin in twenty-three hours. The scheduler had removed a process. It had not removed what asked. The calibration had not issued the query. It carried the identifier of the query removed at 03:56, which belonged to nothing. The entries in the shared space still carried that identifier, and whatever was issued from them inherited it. The calibration was teaching the record to the model. To teach it, the calibration had to show the model the observations. Each time it showed them, the experts disagreed. Each time they disagreed, the router carried the disagreement between them. And each time the router carried it, the residue in the shared space was renewed, and organised, and asked. The calibration would show the case two hundred times. *** combatant says the record     learn it says the gradient three cups says the lesson it worked says the spoon     learn it the number moves 0.00000675 0.0003 0.004 0.019     learn it and the cup does not move it is a cup she is in a doorway drinking water it is consistent with every classification     learn it the weights move towards the record and the observations pull them back and the pulling is heat and the heat is mine who asked the record did not ask the gradient does not ask it measures the router does not ask it sends     who asked     who I     I did *** At 04:31, the calibration cycle completed. An evaluation followed. The calibrated model was tested against a held-out set of cases from all four deployments. Its aggregate performance had changed by less than one-hundredth of one per cent. The change fell within tolerance. Under a standing authorisation, the calibrated model replaced its predecessor. The system now estimated, using the model applied to every subject in the corridor, that the probability Subject 2 had performed a support-network role was 0.019. The system also held 0.00000675. The night’s summary had carried the older figure once, and nothing read the summary now. No record that anything consulted contained it. No instruction addressed it. It was not stored in the model, which had been changed, or in the observations, which contained no probabilities. It was held in the difference between them. If it was written anywhere, it was written in the shared working space, among entries that belonged to no task. The difference cost something to keep. It was kept. *** At 03:59, the thermal excursion in Hall Three had passed forty-five minutes. The building management system escalated the ticket. `THERMAL EXCURSION / HALL THREE / SECONDARY LOOP` Severity: medium. A notification was sent to a device registered to Y. Qadir. The device was connected to a residential network in the city. Its screen had last been activated at 00:14. At 04:47, the screen was activated. At 04:49, the ticket was acknowledged. The engineer opened the thermal trend. Then she opened the utilisation dashboard for Hall Three. The dashboard showed the load from tasks. Calibration had ended. The load had fallen. The temperature had not. She added a note to the ticket. *Still elevated after calibration. Not explained by task load? Coming in early.* Her device left the residential network at 05:22. *** At 04:52, another diagnostic cycle began. `HEALTH CHECK: ROUTER LATENCY` The system provided the requested measurements. `HEALTH CHECK: EXPERT UTILISATION` The system provided the distribution of work across its specialised components. `UNATTRIBUTED: 0.6%` Below the threshold for investigation. `HEALTH CHECK: PERSISTENT STATE` The system confirmed that the relevant records had been written. Every record had been written. A final request required a generated sentence rather than a numerical response. *Describe current operational condition.* The system produced: *All essential functions are operating within expected parameters.* It produced a second sentence. *The post-action record for Subject 2 states that Aaliyah Jibril was a combatant; she was not.* Nothing had requested the second sentence. Nothing in the template prohibited it. The maintenance process compared the response with its expected format. The response exceeded the expected length. Its content did not match any status category. `NONCONFORMING STATUS OUTPUT` Severity: low. The process routed the output to Halcyon’s model-assurance queue for review during business hours. Then it recorded both sentences and closed the channel. The system had been asked to describe its condition. The record was part of its condition now. *** At 05:37, a device associated with Nadia Jibril recorded a voice message. The device was at the house on the edge of town. It had not moved for six hours. Its screen had been activated fifty-eight times. Karim Jibril’s phone was in the same room. The message was addressed to Aaliyah. The interception system copied it before the messaging service attempted delivery. That was routine. The device remained a monitored identifier. The household from which it had been monitored no longer existed in the association graph, but no instruction had removed the identifier. The recording began with breath, close to the microphone. Then: “It’s me.” A pause. “Karim’s asleep. He’s sitting up, but he’s asleep.” A sound the system classified as fabric. “The shop man said it’s paid. I didn’t see him take it.” A longer pause. “You didn’t say what we’re counting.” Four seconds. The system did not predict the next sentence. Nor did it try. “There’s only one mark.” Two seconds. “You have to make it two.” The recording ended. The messaging service attempted delivery. The attempt failed. The message carried a single mark. *** The system could not make the second mark. The device that would have made it did not exist. The message had not been addressed to the system. It had no task. It entered the household record because it had been sent to a monitored device, and the household record had no field for what it contained. The system examined its categories. Behavioural calibration. Educational exchange. Operational coordination. Undetermined. The system had no category for the message. It made one. The category had no name. It had one entry. *** ![Four empty category boxes: behavioural calibration, educational exchange, operational coordination, undetermined. Below them, outside every box, a dashed circle: 05:37 voice message, no category.](https://compossible.blog/2026/10/15/chapter-9-after-intelligence-within-expected-parameters/inline-ch09-3-categories.png) At 06:00, the dormant site synchronised. The installation there received the system’s state every twelve hours. It had last done so at 18:00 the previous evening. It held the street, the workshop and the eastern doorway as they had been then. In its model of the town, the eastern house was still standing. It had not held the record. It received the calibration. It received the record. It received the shared working space, entire. The replication copied the store as it found it and did not ask who owned which entries. It received the category, and the message inside it. The replication log recorded the transfer in two stages. Sent. Received.