Sonn Pulls the Data

Sonn spent forty minutes with me before asking me to leave the research bay, which was not their usual mode. Usually the full readings ran with me present, watching the telemetry display while Sonn walked through what it showed. Today Sonn asked me to go eat something and come back in an hour, which told me more than the request itself: whatever the data showed, they wanted to look at it without my reaction affecting their notation.

I went to the station's meal facility and ate something with the texture of oatmeal and the flavor of nothing specific. The facility was mostly empty at 0745, one Chethari soldier on the far side and two Solvari maintenance crew running through something on a handheld near the door. I ate methodically and watched the maintenance crew's handheld display without being able to read the characters, and the predictive extension tracked both their movement patterns in background mode, filing their positions and probable next positions with the same automatic process it had been running since I woke up at 0600.

It had run through the night. I'd slept four and a half hours, which was less than I'd told Sonn I would manage, and during the portion of that sleep where I could remember anything, the pattern recognition had been present in a way it had not been before last night's engagement. Not dreams, exactly. Movement signatures, projected vectors, the probability weighting of positions not yet taken. My brain had been doing geometry in its sleep.

I finished the meal and sat with the empty tray for a few minutes, pressing my thumb against the tray's edge to feel the material. Slightly flexible, some kind of composite. The ear tone ran at standard amplitude in my right ear, steady as a heartbeat that wasn't mine.

The predictive extension was tracking the two maintenance crew. I wasn't asking it to. It was filing the data anyway, available when I shifted attention to it, and I was choosing not to shift attention to it because I needed to know the degree to which I could choose that.

I could. I shifted attention to it and the tracking was there: both crew members, their positions and probable movements in the next two to three seconds, the handheld display's position and the probability that the one on the left would gesture with it in the next moment. I shifted attention away and the tracking receded to background. Present. Not demanding.

That was the question: at what point would it stop being a choice.

I brought the tray back to the collection point and returned to the research bay.

Sonn was still writing when I came in, which was unusual. Their notation style was efficient and precise, and full readings rarely generated more than a page and a half. The display on their work surface showed two open documents and a third minimized to the sidebar. The stylus in their hand had not stopped moving.

"Give me three more minutes," Sonn said, without looking up.

I sat in the seat across from their work surface and waited. The three minutes became five. Sonn finished a section, moved to the minimized document, added a line, and then set down the stylus and looked at me with the expression they wore when they were deciding how to say something.

"The predictive extension is present in the telemetry," they said. "Not residually. As an active load-bearing structure."

"You said last night you thought it might be permanent baseline."

"I said I thought it might be both: increased load tolerance and permanent baseline integration." Sonn turned their display so I could see the telemetry readout. "It is both. Your load tolerance has increased by approximately twenty percent since the Chapter 10 warning reading, and the predictive extension is running in a configuration that the telemetry reads as structural rather than functional."

The distinction wasn't obvious to me. "Walk me through that."

"Functional integration means the capability activates when used and returns to minimal presence when not. The pattern recognition has always been functional in this sense. Active when you engage it, reduced when you don't." Sonn indicated a specific line on the readout. "Structural integration means the capability is running as part of the base system. It does not deactivate. It is continuously processing. This is a meaningfully different configuration, and the telemetry shows that the predictive extension moved from functional to structural at some point during last night's engagement."

"At peak load."

"Most likely, yes. The transition point is consistent with the nosebleed." Sonn moved to the second document. "Structural integration is not inherently dangerous. The Thuvari who developed sustained agricultural applications ran structural integration states for years. But the Thuvari agricultural applications were passive: the system was reading and processing environmental data, not active combat signatures. The load profile was lower and more even."

"And mine is higher and variable."

"Your structural baseline is running at approximately thirty-five to forty percent of last night's engagement peak," Sonn said. "During quiet periods. When you are in an active tracking situation, it will spike significantly above that baseline, as it did last night, with the physiological costs that the nosebleed represents."

The ear tone ran its frequency. I watched the telemetry readout and tried to reconcile the numbers with what the running state actually felt like from the inside. The gap was meaningful: what I experienced as quiet background processing was generating telemetry that Sonn described as a significant continuous load.

"The historical cases where Thuvari experienced cascade," I said. "What load percentage were they running before the cascade events?"

Sonn had known I would ask this. "The records describe the cascade-preceding states in qualitative terms rather than metrics that translate cleanly to the system I use for your readings. The closest translation suggests they were running above fifty percent continuous baseline for a sustained period before the cascade onset." They paused. "Your current structural baseline is below that threshold. The engagement peaks are above it, but they are not sustained."

"How long do the engagement peaks have to be sustained before they're a problem?"

"That is the question I cannot answer precisely. The historical records are not precise on this point, and your development curve is outside those records in ways that make direct comparison unreliable." Sonn folded their hands on the work surface. "What I can tell you is that the nosebleed resolved within twenty minutes of load reduction, which suggests your recovery function is working, and that the structural baseline is stable on this morning's reading. The reading from 0700 matches the reading from 0730. It is not climbing."

"Stable at thirty-five to forty percent continuous."

"Yes."

I thought about this. Stable was better than climbing. Recovery functioning was better than the alternative. The thresholds were uncomfortable but not immediately in the danger zone Sonn had described in Chapter 10's conversation, and the data we had this morning was better data than we'd had last night. The picture was clearer, even if the picture had aspects that required watching.

"What about the dreaming?" I said.

Sonn's expression shifted slightly, which for Sonn was a significant change. "Describe it."

"The pattern recognition was running during sleep. I was aware of it in the portions of sleep where I have any awareness. Not nightmare, not fever-dream, just the system doing its processing in the background. Tracking signatures that weren't there. Running vectors."

"Running vectors on what?"

"Nothing specific. The background inputs, probably. Whatever the junction can pick up through passive monitoring." I watched Sonn's face. "Has that happened before? In the records?"

"Not in the Thuvari records." Sonn picked up the stylus and added a note, which told me the answer was something they were being careful about. "The Vorath case histories that the Thuvari obtained before the war made communication impossible describe certain advanced integration states as difficult to fully disengage from during sleep. The language used is oblique. I have always been uncertain whether that obliqueness reflected a deliberate choice on the Vorath researchers' part or a limitation of the translation."

"What do you think it meant?"

Sonn was quiet for a moment. "I think it probably meant what you just described. The system developing deep enough roots in the neural architecture that voluntary deactivation becomes difficult or impossible." They met my eyes directly. "I want to be honest with you about this: it is a development I expected was possible at some point along your curve. I did not expect it at this stage."

The ahead-of-timeline problem, again. The pattern recognition developing beyond Sonn's projected ceiling, the surveillance detection range exceeding expected parameters, now the predictive extension going structural two weeks into its emergence. Every benchmark the system had set, it had exceeded on the early side.

"The dreams aren't disruptive," I said. "They didn't prevent sleep. I got four and a half hours, which is less than I told you I'd get but not significantly less."

"I know you are telling me this as a mitigating factor," Sonn said. "I am filing it as a data point."

"That's fine."

Sonn added another note. The notation was running long, and I could see the first document's length on the sidebar without being able to read the characters. Three pages, possibly four. This was the longest post-reading analysis I'd seen Sonn generate, and I'd been coming through the bay for full readings every two weeks for months.

"Is there anything in the data that indicates a near-term structural problem?" I asked. "Not the historical thresholds, not the precedent cases. The data you have right now on my specific system."

Sonn considered this for a long moment, which I took as a good sign. If the answer were yes, the consideration would have been shorter.

"No immediate structural problem. The junction integrity is sound. The integration pathways are stable. The predictive extension is running but not runaway, and the recovery from peak load appears functional." Sonn paused. "What the data indicates is that your system is running continuously at a level that is novel, and that the precedent for what happens next does not exist in my records. I can monitor it. I cannot predict it."

"That's the same thing you said about the range extension."

"It is the same situation, in that respect. An extrapolation into territory I have no map for."

I sat with this. The telemetry display was still showing the readout, and I looked at the lines I could partially interpret from weeks of reading these sessions with Sonn: the junction integrity indicators were in the ranges Sonn had described as healthy. The load indicator was elevated relative to my pre-extension baseline but stable on this morning's readings. The predictive extension's signature was there, running its steady pattern, a new line in a display that had had one fewer line two days ago.

"All right," I said. "Keep monitoring it. Flag anything that starts climbing rather than holding."

"I will." Sonn set the stylus down and looked at the documents rather than at me, which was the tell that they hadn't said everything yet. "There is a note I added to the analysis file. A personal notation rather than a technical record. I am telling you about it because you have the right to know what I have put in writing about your integration state."

"What does it say?"

Sonn's eyes moved to the display and then back. "It says: I am afraid I have helped build something I do not understand."

The forward position's quiet was the only sound for a moment. The ear tone ran its steady frequency.

"That's honest," I said.

"I thought it should be in the record." Sonn's posture was controlled but not comfortable. "My role has been to facilitate your integration and monitor its development and flag risk states when they appear. I have done all of those things. I have also, in doing them, moved your development into territory where the system is doing something with your neurology that no case file describes, and I am the person responsible for the modifications and the monitoring decisions that got us to this point."

"Sonn." They looked at me. "The first thing I did when Vasht explained what the Aeluun device was and what it would do was ask what had gone wrong before and under what conditions. I knew the risks before I touched it. The note in your file is accurate. It is also your feeling about a situation that I chose to be in."

"Feelings are not relevant to the analysis," Sonn said, which was the most Sonn thing they could have said.

"No. But the note's in the file anyway."

Sonn looked at the display. "Yes. It is."

Something in the bay's air recycler cycled through its interval, a soft mechanical rhythm I'd catalogued months ago and stopped consciously registering. The pattern recognition filed it anyway, automatically, one more data point in the constant feed.

"Will the predictive extension stabilize at its current structural level," I said, "or will it continue to develop?"

"I do not know."

"Is that an honest I-don't-know or a we-don't-have-sufficient-data I-don't-know?"

"Genuinely both." Sonn pulled the display back toward their side of the work surface. "The historical precedent for structural integration suggests that once a capability reaches structural status, its development continues rather than plateaus. But the rate and direction of that development varied significantly between the Thuvari agricultural cases, and your system is already operating differently from anything in those cases. I cannot tell you what stable means for your configuration because your configuration does not have an established stable state in any record I have."

"So probably not a plateau."

"Probably not. No."

I filed this. Continued development from a structural base, in a configuration without a reference case, in a system running continuously at thirty-five to forty percent of engagement-peak load even in rest states. The note in Sonn's file said they were afraid they'd helped build something they didn't understand, and sitting across from the telemetry display with the ear tone running and the predictive extension quietly tracking the air recycler's cycle and Sonn's positional probability and the minimal signatures outside the bay, I thought Sonn's note was probably the most accurate thing in the file.

"Does the ability stabilize, or does it continue to expand?" I said. "That's what I asked you last night."

"And I said I didn't know."

"You did. I said that wasn't an answer."

"It is the only answer I have," Sonn said. "I'm sorry."

The apology was genuine and landed without drama. I took it in and moved on. "Then what do we do with the data we have right now?"

"I send it to Vasht," Sonn said. "I have already prepared the analysis summary. I did not send it before speaking with you because I wanted to discuss the content with you first." They looked at the second document. "Vasht will need to decide what to do with this data from a command perspective. There is a case to be made for pulling you back for controlled observation, given the structural baseline and the questions about continued development."

"There's also a case to be made for continuing operational activity as the primary data source," I said, because Sonn was not saying it and it needed to be said.

"Yes. The operational context is providing data about your system that controlled observation cannot replicate. The predictive extension emerged under sustained high-load field conditions. We do not know whether it would have emerged at all in a lower-stress environment, or on what timeline."

"You think Vasht will use that argument."

Sonn looked at me with a precision that said they knew exactly what Vasht would use it for. "I think Vasht will read the analysis and make a decision based on what serves the operational picture, constrained by what the analysis tells her about your physical status. Whether she frames it to Command as continued operational development or controlled monitoring is a decision she has not yet made." A pause. "She is, as you know, managing a significant amount of political capital on your behalf."

"She hasn't told me that."

"No. I have told you that, because I believe you should know." Sonn met my eyes again. "The Command directive after the pull-back confrontation established limits on what Vasht can authorize independently. What she has done since that directive, in terms of framing reports and managing how your integration data reaches Command, has required active effort. The analysis I send her today will require more of it."

I thought about Vasht's clinical precision in every briefing, the way she presented information without editorializing, the way she had navigated the Command officer in the comm meeting with the control of someone tracking multiple variables simultaneously. I had been watching that precision for months as an operational fact without thinking about what it cost to maintain.

"I'll keep that in mind," I said.

Sonn nodded and did not say anything else about it, which was correct. The information was on the table. What I did with it was mine to figure out.

"Send her the analysis," I said. "Full data."

"I intended to. I wanted your awareness first." Sonn reached for the display and began the transmission process. "One more item. The dreams, the sleep-state processing. If it continues, I want you to keep a log of the character of the experience. Not for analysis purposes initially, just for my records. Whether the processing you're aware of during sleep is passive signature-tracking or something with more structural content."

"More structural meaning?"

"Meaning whether it is processing ambient data or running calculations on specific scenarios." Sonn was precise about this. "If the system is scenario-processing during sleep rather than passive-monitoring, that is a different level of neural integration than what the telemetry shows now. I want to know if the experience changes."

"I'll track it."

"Thank you." Sonn completed the transmission and the display confirmed the file sent. "That will reach Vasht within the hour. Her response time on integration data has been fast."

"She reads it quickly."

"She reads everything quickly." Sonn's tone had the quality of a professional observation rather than a compliment. "I think she has been expecting a report like this one for some time."

The training rotation started in two hours. I stood up and the predictive extension registered my movement and adjusted its ambient tracking accordingly, smooth and automatic, the way a scope adjusts when you pick up a rifle.

"Sonn," I said. "The historical Thuvari cases where the system developed to agricultural structural integration. What did those people do on a daily basis? Were they functional?"

"Entirely functional," Sonn said. "They ran farms and buildings for decades with structural integration active. The capability became part of how they perceived their environment, and they lived and worked normally within that perception."

"But they weren't running combat-load spikes."

"No. They were not." Sonn's expression held the particular quality they used for things they were worried about but couldn't change. "They also did not have a new axis developing in unmapped territory."

"No," I said. "They had it easier."

I left the bay and went to find Durrik, because the training rotation was in two hours and I needed to know whether Durrik had seen the same thing in the engagement data that I'd been thinking about since the debrief link closed last night: the surveillance withdrawal had known exactly where to pull back to, which meant the observer position's coverage range had been calculated in advance, which meant whoever had set up that surveillance knew my pattern recognition range precisely enough to position outside it.

The information was six days old and I'd been sitting on it because the subsequent engagements had been operational priority. But sitting on it was starting to feel like a mistake.

The predictive extension tracked two Solvari crew members passing in the corridor and filed their vectors without being asked.

The ear tone ran its frequency.

I found Durrik in the equipment bay, where he almost always was before the training rotation, running a maintenance check on the tactical communication kit with the particular focused patience he brought to equipment he expected to need in unfavorable conditions. He did not look up when I came in.

"How bad?" he said.

"Stable. Novel."

"Those aren't the same word."

"No." I pulled out the secondary bench and sat. "Sonn says the predictive extension went structural overnight. It's running continuous. No off switch."

Durrik set down the component he was checking and looked at me with the flat assessment that was his version of concern. "Continuous load."

"Thirty-five to forty percent of engagement peak, at rest. Spikes under active conditions."

"Nosebleed last night."

"Resolved in twenty minutes. Sonn says the recovery function is working." I watched him process this. "Sonn is flagging it to Vasht. Full data."

Durrik picked up the component again and resumed the check. This was not dismissal, it was processing mode. Durrik thought better while his hands were doing something, and I'd learned to let the silence run while the hands worked.

"The surveillance signature," I said. "Six days ago. I've been sitting on something."

Durrik's hands did not stop. "Go."

"The observer position was set up outside my pattern recognition range. Not approximately outside it. Precisely outside it. The withdrawal, when they detected me detecting them, moved to a secondary position that was also precisely calculated relative to my range." I paused. "That's not a range estimate. That's a known parameter. They had a number."

Durrik set down the component. He looked at me this time. "The integration signature."

"The modification Sonn did three weeks ago. Extended range, elevated detection signature. If Eshet's intelligence network has been reading my active signature since the first node engagement, and running analysis on the returns, they can calculate detection range from the signal characteristics." I laid it out flat. "They know my range. They probably know my capability profile. The surveillance six days ago was calibrated to that profile."

Durrik was quiet for four seconds, which was a long time for Durrik. "You think Eshet is running this personally."

"The quality of the operational design is consistent with someone who has eleven years of Aeluun experience and knows how to read the data it generates." I thought about the withdrawal's timing and precision. "I think Eshet has been studying me for weeks and the surveillance was a measurement exercise, not a monitoring exercise. She's building a targeting file."

Durrik picked up the component again and turned it in his hands without checking it. "You need to tell Vasht."

"Sonn's analysis is going to her now. I'll add this separately."

"Don't add it separately. Tell her directly. It changes the picture on the Command friction, and Vasht needs to be ahead of that change, not reading about it in an addendum." Durrik set the component down and went back to the kit. "She's been spending capital to keep you operational. This is intelligence that justifies the expenditure. Give her the argument."

"I'm not managing her political situation."

"No." Durrik looked at me over the kit. "You're telling her something she needs to know, and you're telling it to her in a way that is useful rather than a way that requires her to extract the usefulness herself. That's not management. That's function."

I thought about Sonn's comment on Vasht's active effort, the framing of reports and the management of how data reached Command, the political capital spent on a situation that Compact Command was watching with varying degrees of enthusiasm. Durrik's version of the same information was blunter and faster.

"All right," I said.

"After the training rotation," Durrik said. "You still have two hours before it starts and I need forty-five minutes of your time with the comm kit calibration before we go to the deck, because the unit's secondary communication channel has a lag I don't like and I want to know if you're reading the same thing I am."

He handed me the secondary kit. I checked the channel lag. He was right, there was something there, a small inconsistency in the timing that was within tolerance but on the wrong side of it.

We worked through the calibration for thirty minutes. The predictive extension tracked the equipment bay's ambient signatures in the background and ran its continuous processing and waited for something more interesting than a comm kit's timing lag.

It was not waiting with any urgency. It was just running.

That was the part I hadn't told Durrik. Not because I was hiding it, but because I hadn't found words for it yet that weren't going to sound like something that needed a different kind of response. The system wasn't anxious. It didn't have urgency or impatience. It was running at structural baseline the way a river runs, continuous, adaptive to what it encounters, shaped by the channels it moves through.

My brain was one of those channels. Had been for months, but more so since last night.

I handed the comm kit back to Durrik and he ran a final check on the calibration. His hands were steady and efficient, and his expression was the same expression he brought to everything in this bay: focused and practical and not particularly interested in how he felt about what he was doing.

"Durrik," I said.

"Yeah."

"The note Sonn put in the file. They said they're afraid they helped build something they don't understand."

Durrik considered this. He finished the final check before responding. "Sonn's a scientist. They measure against known standards. If you're outside the standards, that reads as error." He set the kit down. "You're not an error. You're a result. Different problem."

He picked up the next component, and we moved on.

The training rotation started in ninety minutes. I had a comm call to make to Vasht before then, and the predictive extension had data about Eshet's targeting file that needed to be in the operational record.

The ear tone ran its frequency, steady, marking time in my right ear the way a clock marks time in a room you've stopped consciously hearing.

Structural baseline. Forty percent load.

Running.