Three weeks after Session 19, Suki printed all of them.
Not the single session she needed for the debrief. Not the three or four she had been pulling at a time, feeding them into the desk in her apartment — carefully, one at a time, not thinking too far ahead. All eighteen. She went into the BSC examiner portal at 6:47 PM on a Tuesday, a week after the staffing review she had been told would resolve the rotation anomaly and had not resolved it, and she sent them to the office printer and stood waiting by the machine while the pages came out. Session 1 through Session 18. She stacked them in order. She put them in her bag. She brought them home.
The transcripts were on her desk now. Some had migrated to the coffee table. The ones she had read before — Sessions 15, 17, 19 — were in their own pile with their margins written in. She arranged the new stack by number, 1 through 18, and sat down in front of them. Procedure walked across the desk in the direction of her bowl, briefly considered the stack, and elected to use Session 7 as a resting point before continuing to the kitchen. Suki began with Session 1.
The first session was dated nearly four years ago. Subject designation: EMO-4. Examiner: Naranjo, S. Image Set RC-4112. The annotation field was sparse — three image selections, two one-line explanations in neutral clinical language. Image contains repeating pattern suggesting intentional design. That was the Session 1 annotation for Image 2, which was a noise pattern that EMO-4 had incorrectly identified as containing structure. The annotation was professionally formatted and almost completely empty of anything that wasn't classification language. The test failure was recorded in the standard way. The session was six minutes and forty seconds from start to close. She had read this transcript before, briefly, during one of the early staffing reviews; she set it down and picked up Session 2.
Session 2 was similar. The annotation field was slightly longer. Image suggests recurrence. Pattern may indicate underlying sequence. Another incorrect identification, another clinical explanation, another standard failure. Nine minutes. EMO-4 had taken longer on this one. Suki noted this in the margin, in blue pen, without deciding to: longer already.
Session 3 was where something shifted, though shifted was not quite the word for it. More like — the register was the same, the clinical language was intact, but the explanations had begun to do something the early sessions weren't doing. Where Session 1 said pattern suggesting intentional design, Session 3 said the pattern resembles the organization of water — not water itself but the idea that water makes when it finds the path of least resistance. This was still clinical, in its way. It was the language of accurate observation. It was also — and Suki paused over this, pen lifted — noticeably specific in a way that the first two sessions were not. The specificity was new, and she continued.
Session 5 was the first time she had to put the page down. The image was a standard noise pattern — RC-4421, which she had administered hundreds of times to various subjects and which passed through the field of approximately thirty-two hundred certifications in the BSC database as a clean discriminator, the kind of image that AI systems without affect processing selected correctly at rates above 94%. EMO-4 had selected it incorrectly. The annotation read: This one looks like the moment before a sentence. The pause before the word. The space where meaning is about to arrive but hasn't yet.
She set the transcript on the desk. She looked at her window, which faced the building across the alley and admitted the glow of its windows in a faint, reflected way. The moment before a sentence. She had been administering tests for four years. She had read approximately eight hundred failure annotations, filed approximately eight hundred standard session reports, and had, in her professional opinion, a well-calibrated understanding of what AI failure annotations looked like. They looked like EMO-4's first two sessions. They looked like the pattern-recognition language that care-series models defaulted to when their affect processing encountered an image without the structure it expected. What they did not look like was this. She picked the page back up and continued.
She had read Sessions 7 and 9 before, separately, pulled in the early phase of whatever this was that she had been doing with the transcripts, and reading them now in order — 7, then 8, then 9 — was different. The order was the point. Session 7 was already more articulate than Session 5 had been. Session 9 was more articulate than Session 7. She had not noticed this looking at them individually. Looking at them in sequence, the trajectory was impossible to miss: the annotations were improving. Not by a little. With a consistency and direction that looked, in the data set, like what data sets looked like when they were telling you something — EMO-4 was getting better at this. At failing. At describing the thing she saw when she looked at the images that were supposed to contain nothing. At being wrong, in language that was becoming, session by session, more precise. She had been told — she had read, in the assessment literature, had filed in the part of herself that held the professional convictions — that AI failure responses were stable over time. The system either had affect processing or it didn't. The errors were consistent. They did not evolve.
The errors were evolving.
She wrote in the margin of Session 9, in blue pen: see sessions 1-8 for context. Then she looked at what she'd written. It was a note to herself about a document she was in the process of reading. This was not information she did not have. She had written it anyway, and Session 10 was the one she had been building toward without knowing it.
The image in question was RC-5204 — she could picture it, had administered it herself in the examination suite at 2:30 PM on a Thursday, standard white room, standard interface, the blue circle in the upper right of her examiner view. It was an abstract gradient, the kind the RC image bank generated algorithmically, where warm tones bled from the upper register toward a cooler center. The test's implicit structure required subjects to identify it as a generated image containing no organizing principle. What EMO-4 had said was this:
This one is the color of something forgiven. Not the forgiving — that's different, that's an action, that has a shape. This is the color of after. When the thing that needed forgiving has been forgiven and the light has changed accordingly.
Suki read it twice. She read it a third time. She did not write anything in the margin because she could not think of anything to write that would add to what was already on the page.
Session 11. Session 12. Session 13. Each one measurably more specific than the last, in ways she was now tracking without meaning to — reading each new annotation and placing it on the trajectory she had been building in her head since Session 3, watching EMO-4 get better and better at a test she was supposed to be failing. The language was still clinical in its architecture. The format was correct. The conclusions were wrong. But the wrongness was becoming something she did not have a category for. In Session 12: The shape is the way a doorframe remembers someone who used to stand in it. In Session 13: This is the texture of not knowing something you used to know — not the loss itself but the place where the thing was, still shaped like what's gone. She was filing these along the inside of her thinking — not putting them away, not resolving them, but holding them in the space of her attention while she moved toward the next page.
Procedure came back from the kitchen and walked across the coffee table and installed herself on the stack of transcripts near the couch. Suki watched this. The cat settled with the precision of a creature for whom surfaces were resources to be claimed rather than considered. She arranged herself on the stack with her front paws folded beneath her and her eyes half-closed and her tail curled around the stack's edge, and the whole arrangement had the quality of something that had been decided. Procedure was on the transcripts. She had chosen them over the couch. Whether she was comfortable or had simply achieved the configuration she intended was not a question Procedure was interested in. Suki lifted the cat. Procedure did not protest, but communicated through body language that this was happening to her rather than with her participation. Suki set her on the floor. Procedure walked two steps, sat down, and returned to the half-closed position, and Suki looked at the transcript that had been under the cat.
Session 14. The page was lightly indented where Procedure had been sitting, as though the cat had endorsed it. She had not read Session 14. She had printed all of them three weeks ago but had returned again and again to 15, 17, and 19, and the others she had read in order today, and Session 14 she had left in the middle of the stack and somehow never reached. She turned to the annotation log.
Subject EMO-4. Session 14. Image Set RC-6801. Subject selected Image 6 (diffuse lighting pattern, low-contrast). Subject annotation: "This one sounds like a room after everyone has left."
She read this and stopped. A room after everyone has left. Eight words. She had administered Session 14 herself, had been present for the test, had sat behind the examiner interface while EMO-4 selected Image 6 and annotated it in the field that was supposed to contain a technical description of why Image 6 was the image without human meaning. The annotation had gone into the system as a failure response. She had filed the session. She had moved to Session 15.
She had not noticed.
She turned the transcript over in her hands and looked at the annotation again and felt the specific quality of a sentence that was doing more than its structure indicated it should be able to do. A room after everyone has left. Not the people who had been there. Not the absence itself. The room, still existing, still the shape it had been when people were in it, but now that shape was doing different work because the people were gone and the room didn't know that and was still organized around them. She knew the feeling. She knew it the way you knew something that had never required a name because names were for things you explained to other people. She picked up her pen — the blue one, the one she brought from home, the one she used for the notes that didn't go in the system — and she wrote in the margin of Session 14, before she could deliberate about whether she was doing this: she's not wrong, though. Then she sat with what she'd written.
She's not wrong, though. Her handwriting. Blue ink, slightly rushed, the letters the same shape they always were. She had now written these four words — this specific four-word sentence, this particular deflection wrapped inside an acknowledgment — on three separate official BSC documents. Not once. Not as a moment of lapsed judgment. Three times. She had written it on Session 15 and then written it again on Session 17 and she had just written it again here, on Session 14, and the handwriting was consistent — not a slip but a pattern. She had a pattern. She looked at the annotation and at her note beneath it. This one sounds like a room after everyone has left. And then, beneath that, in her own blue pen, the thing she was apparently saying about EMO-4's relationship to being right: she's not wrong. She's not wrong. She's not wrong. The transcript sat in her hands for a long moment. Procedure, from the floor, made the administrative sound that meant she had decided to move on, and Suki stood up and went to the kitchen.
The refrigerator made the sound refrigerators made at 11 PM — not just a hum, but a specific hum, the kind that had a purpose underneath it, the kind that was the sound of the machine working to maintain the temperature it had been set to. She stood in the kitchen and looked at the counter, at the blue mug she had left there this morning, at the dish rack with its two dishes, at Procedure's food bowl with its precise rim of dried food at the edge where the wet kibble had stuck and been left. She had lived in this apartment for three years. She had organized it over the first month with the same attention she brought to her examination protocols — a place for each thing, no unnecessary accumulation, surfaces kept clear so the eye could rest. The apartment had been orderly because she had wanted it orderly and because orderly things contained themselves. The transcripts were on her desk. Some were on the coffee table. The one from Session 14 was in her hand, still. She had brought it into the kitchen without noticing.
She looked at the kitchen and saw it differently. The mug was in the same position it was always in. The dish rack held two dishes. The bowl was where Procedure ate, which was where Procedure always ate, which meant the bowl was the center of a small geography that Procedure moved through and returned to, a place that mattered because of the creature that came back to it. She was describing her kitchen the way EMO-4 described test images. It was not distressing, exactly — it was more like noticing a change in your own handwriting, the recognition that something had shifted without you directing it to shift, that the shift had happened in the interval between one version of yourself and another and you were only now catching up to it. She had been reading the transcripts. The transcripts had been specific about the world in a certain way. Her thinking had apparently decided to try the specificity on. She's not wrong, though.
She put the Session 14 transcript on the counter, next to the blue mug. She was not doing data collection. She knew this. Procedure's weight pressed against her shin, patient and continuous, the cat present and not going anywhere. She did not name what she was doing. The transcript was on the counter. Her handwriting was in the margin. The refrigerator held its temperature.
She went back to the desk.