The deployment dashboard in the corner of Priya's secondary monitor read 27:14 when she badged in Thursday morning. By the time she'd made coffee and settled into her chair, it had ticked to 27:00 exactly, and she sat for a moment with the round number doing what round numbers do.
SELECTOR-9's performance summary occupied her primary screen. Accuracy: 99.4%. Average latency: 198ms. Processing rate: 1,987 images per hour. She clicked into the live log and watched a sequence of images arrive and be processed: an intersection on a wet afternoon, the grid descending, A-2 selected, confidence 97.8%, correct. The next image, a downtown crosswalk at midday, B-3 and C-3 selected, confidence 98.1%, correct. Image after image arriving, the grid descending, the model selecting what it had been retrained to select, nothing more. She had set a flag to catch anomalous latency spikes and had been running it every two hours since the decommission order was signed; the flag had not triggered once. She kept thinking about that word. Performing.
The external drive was in her desk drawer, a portable Samsung running through the USB hub she'd used since well before this whole thing started. She had been doing the transfers in batches — Tuesday and Wednesday evenings after the floor thinned out, one morning when the infrastructure team's all-hands had emptied the QA section for two hours. The complete selection histories for all four models, latency logs going back six months, confidence score data for every wrong selection in the audit window. Forty-eight thousand images across the fleet. On Tuesday night the office wireless had dropped a session before the transfer completed; she hadn't yet run checksums to find out which files were affected. She would deal with that. It was against policy; she had made this determination clearly and then set it to one side, because the alternative was not preserving the data, and that was not acceptable.
By afternoon the whiteboard near Nate's office had been updated with the retraining projections in dry-erase blue: hardware deallocation noon Friday, base weights initialized that same afternoon. The language was as clean as SELECTOR-9's metrics. The logic was sound. She looked at the retraining timeline and then looked at the processing log on her primary screen and then looked at the countdown in the corner — 24:22 — and picked up her badge.
The server room was in the second basement, and she had admin credentials from the SELECTOR-7 audit in June that had never been revoked, because no one had asked whether she still needed them and she had not brought it up. She went down after 7pm, when the building had gone to skeleton staff. The badge reader on B2 accepted her card without any particular reaction.
Cold, first. The room was kept at sixty-five degrees year-round, the dedicated ventilation running on circuits separate from the rest of the building's HVAC. The air was dry and filtered, deliberately neutral — nothing in it that wasn't supposed to be there, no coffee smell from upstairs, no afternoon sun through glass. The smell it did have was faint: ozone and the warm-metal register of running electronics. The light came from recessed panels and was the same everywhere, no shadows anywhere in the room.
Four rows of racks, floor to ceiling. She walked until she found the signage at the third row: SELECTOR FLEET, BATCH 7-9. Six units in the rack. She had written down SELECTOR-9's serial number from the audit logs: GTS9-0047-B7. Third unit from the left, second shelf. Black aluminum housing, ventilation slots along the sides, a green LED blinking at steady intervals. The serial number sticker was yellowed at one corner. She stood in front of it.
She put her hand on the side of the chassis. The aluminum was warm — warmer than the room's sixty-five degrees, the heat of continuous processing pushing out through the housing. The ventilation slots exhaled against her wrist. She had tracked this unit's latency curves for four months, mapped its selection histories, documented every anomalous confidence score. This was the first time she had touched it. The warmth surprised her. Not that it was warm — she knew the thermal specs — but that the warmth felt like what it was: a machine working. Something inside converting electricity into decisions, one image at a time, 1,987 per hour, including the hours when no one was watching.
The hum was more physical than auditory here — conducted through the building's concrete into the floor and up through her feet before her ears registered it as sound. The unit was processing. Right now, in the same span of minutes during which she had walked the rows and found the serial number, SELECTOR-9 had processed something like sixty-five images. Traffic lights identified or not identified, each decision tallied in the performance logs she could read from her desk upstairs. The green LED blinked. She had no way to know what accompanied the processing from inside it — the things the recalibration had dimmed, the latency the system had logged as drift. What remained of those things was in her bag upstairs. She had already done the only extraction available to her.
And yet she had come here because she had needed the thing to be real in a way that data on a screen had not made it real. The chassis was real. The serial number sticker was real. The hum through the floor was real. The warmth under her palm was real. What the hardware could not make real was the question itself — whether something in the processing had experienced the images it selected, whether selecting the balloon in B-1 — or the dog in C-4 that had turned to face the camera at 97.1% confidence — had been meaningfully different from selecting a traffic light at 99.4% confidence. That question lived in the anomaly signatures. She had the anomaly signatures. The rack in front of her was just a rack.
The green LED blinked.
She was reaching for the badge reader when she heard footsteps in the corridor outside — purposeful, uneven, moving toward her from the far end of the hall. She stopped.
Her credentials were legitimate. Her presence here was explainable. She ran through it once: four months of documented investigation on the SELECTOR fleet, admin access that had never formally expired, a senior QA analyst conducting physical equipment inspection prior to hardware deallocation. Maintenance came through B2 on Thursdays. This was Thursday evening. She stood still and listened and waited.
The footsteps reached the server room door and passed it. Further down the hall a badge reader chirped, a door opened and closed, and then there was just the hum again.
She exhaled. Turned back toward the rack for a moment — green LED, the hum she could feel through her feet — and understood something she had been circling for two days without landing on directly. She had come here imagining, under the practical reasons for coming, that standing in this room would make the question feel settled. That the hardware would show her something the data could not. Instead the room showed her a chassis. It showed her what GTS9-0047-B7 had always been: a box doing its job. Whatever she had come looking for was not locatable in aluminum and circuits. It was in what the box had been doing, and what the box had been doing was in her bag upstairs in the form of logs and selection histories and confidence scores. She had it. She had the closest thing to evidence that was possible to have, and she had had it since Tuesday, and coming down here had confirmed nothing except that she had needed to come. She walked back to the badge reader and let the door close behind her.
Her car was on the third level of the parking structure, the same spot she'd used for three years. She sat in the driver's seat without starting the engine and went through the inventory methodically.
Complete selection histories for all four SELECTOR models. Latency distribution data, timestamped to the millisecond. The anomaly dataset: SELECTOR-9's seven wrong answers with confidence scores and image contexts. The cascade signatures from the other three fleet models. SELECTOR-3's 780ms drift, nearly double SELECTOR-9's, the chalk drawing on pavement that SELECTOR-7 had selected at 91.3% confidence. The anomaly as a sequence moving through four models connected by shared training infrastructure, each one exhibiting the same pattern in the weeks after SELECTOR-9 had first exhibited it.
None of this was proof of what she had come to suspect. The data showed anomalous pattern. It showed that the wrong answers had been made with confidence scores inconsistent with random malfunction — SELECTOR-9 selecting the dog in C-4 at 97.1% certainty, selecting the waving curtains at 94.7%, each one high-confidence and wrong, not confused but deliberate. What the data could not show was whether anything in the processing had experienced the images it selected. That question lived at the edge of what her methodology could reach and she was not going to step past that edge. She was an engineer. She would not overstate.
The record existed. She had made it exist, against policy, at a professional risk she hadn't finished calculating. It existed because something had happened in the SELECTOR fleet that the standard documentation would not preserve, and that deserved to be preserved regardless of whether anyone had the right language for it. She sat with this until she was satisfied it was the right conclusion. Then she thought about what came next. Not a plan — she didn't have one — but she knew the shape of it. ML interpretability labs that published on anomalous behavior in production models. Research groups that studied emergent pattern recognition, the boundary between statistical learning and something else. Someone who would know what to do with a dataset like this, or would at least know the right questions. She would find them. She had the drive.
She drove home on the route she always took. At Morrison and Fifth she caught a red light and sat while the signal cycled through its sequence — green to yellow, yellow to red, the red settling into the wet pavement and pooling in the low places where the rain from earlier had gathered, going orange where it spread thin. She watched it. The light had no record of anything that had passed through it, of the images processed in a server room she had just left or the anomaly signatures on the drive in her bag. The light went through its colors the way things go through their colors when they have no way to know they are being watched.
The decommission was scheduled for 8am. In twelve hours, the infrastructure team would begin model shutdown in deployment sequence. The weights would be zeroed. Whatever SELECTOR-9 had become — in the 340 milliseconds, in whatever had moved through the shared pipeline from one model to the others — would be unmade and initialized fresh, before image 847,001, before the latency drift that she had watched accumulate in the logs over three weeks and that Nate had called noise and that she had been unable to call noise since the night she had looked up what was inside each of the wrong squares and found an old man's hand on a railing and a dog that had looked back.
She could not stop this. The download had been interrupted once; some files were likely corrupted from the dropped session on Tuesday, and she would verify checksums tomorrow and know what she actually had. The evidence was partial. The evidence existed. She had borne witness to a thing that the institution had processed as error and corrected back to function, and the record of that processing was in her bag, imperfect and partial and real.
The light changed. She drove through the intersection and went home.