The floor was empty by 7:30. The other three analysts had left for dinner and whatever came after dinner -- Priya had heard them debate a bar downtown, had said she might catch up later, which meant she wouldn't. She had the selection history open in front of her: a 12,000-row CSV pulled from SELECTOR-9's processing logs, each row an image reference, each image a decision. The latency numbers were in column F. She wasn't looking at column F.
She was looking at column H, which tracked per-image object confidence. Not selection confidence -- that was column G, and column G was fine, the expected distribution with expected variance, nothing to flag. Column H was what SELECTOR-9 returned when asked how confident it was that the selected square actually contained a traffic light. For most images, column G and column H said nearly the same thing: high selection confidence, high object match. The model selected because it was sure; it was sure because the traffic light was there.
She'd been working on a script since 6 p.m. to find cases where those two numbers diverged -- where SELECTOR-9 had selected a square with high confidence but logged low traffic light match for the same square. The logic was straightforward: if the model selected something and was certain about the selection but couldn't find the object it was supposed to be selecting, what was it selecting?
The script ran in ninety seconds and returned seven rows. She checked her query for errors, found none, ran it again. Seven rows. She opened the output: seven images, dates ranging across the past three weeks, each with the same pattern -- selection confidence ranging from 94.3% to 97.1%, traffic light match confidence ranging from 0.2% to 0.8%. Below noise threshold. The model had been completely certain it was selecting something. It had been nearly certain that something was not a traffic light. Not a confused model. A model selecting on different criteria.
Outside, through the glass wall of the bullpen, the GridTrust campus was lit up and quiet, the parking lot half-empty, the distant server building running its usual LEDs. The kombucha tap in the break room gurgled once and fell silent. Priya's monitors cast her face in blue. She double-clicked the first image file.
C-2, confidence 94.3%. The image was a suburban intersection, late afternoon, the light on the pavement going flat with the end of the day. The traffic light was in the upper left corner of the frame, clearly in A-1 -- a standard four-way signal, red lens lit, unmistakable. And then her eye moved to C-2, the square SELECTOR-9 had selected alongside the correct one. A woman mid-stride on the sidewalk, one arm crooked under a grocery bag that had shifted, her elbow compensating for the new weight distribution. Nothing in C-2 was a traffic light. The woman had been going somewhere and the bag had slipped and she had adjusted, and SELECTOR-9 had marked that square at 94.3%. The image resolution was high enough that Priya could see the bag was paper, not plastic -- the creased top folding under the woman's grip.
B-1, confidence 96.1%. Urban intersection, morning, power lines against a pale sky. The traffic light was in D-3. In B-1, a red balloon tangled in a wire above the crosswalk, the string gone slack, bright round red cut loose from whoever had held it.
C-3, confidence 96.2%. Downtown, overcast. The traffic light in A-2 fully visible. In C-3, a pigeon frozen by the shutter in the moment between going up and coming down, both wings extended, the breast feathers ruffled against the grey sky.
D-4, confidence 94.8%. A residential intersection, dry afternoon, old concrete. The traffic light in B-2 was standard. D-4 was a sidewalk corner where the pavement had fractured along a branching system, cracks spreading from a main split outward in irregular branches.
D-2, confidence 95.4%. Bridge intersection, fog. Traffic light in A-3. At the right edge of the frame, an old man's hand on a metal railing -- only the hand visible, the rest of him outside the frame, knuckles raised and white at the grip points.
C-4, confidence 97.1%. Commercial intersection, midday. Traffic light in B-1. A dog waiting at the curb on a leash, its head turned to face the camera directly. Both eyes open, aimed at the lens.
A-4, confidence 94.7%. Residential intersection, windy day. Traffic light in D-1. A second-floor window fully open, white curtains billowing out through the frame into the exterior air.
Priya sat back in her chair. Seven images, seven squares containing no traffic lights. Selection confidence ranging from 94.3% to 97.1% -- not the confidence of a model that was uncertain or degraded, but the confidence of a model that knew exactly what it was selecting. She looked at what it had selected: a woman balancing groceries. A balloon loose from its string. A pigeon in mid-flight. A crack in concrete. An old man's grip on metal in fog. A dog's eyes aimed at the camera. Curtains past the window they were made for.
She had the analyst's instinct to name the category before she let herself think about what it meant. She looked for the pattern. Things in motion -- the pigeon, the curtains. Things in contact -- the hand on the railing, the woman's arm shifting for the bag. Things with color that read like something else -- the red balloon that was not a signal. Things that had no function in the frame, that were simply there.
She couldn't make the category cohere. The dog was not in motion and not in contact and not the wrong color. The crack in the sidewalk was not alive. Whatever connected them was not in any taxonomy she could map onto the data. She had been looking at these seven images for eleven minutes, the monitor glow turning her notes blue, and she got up to make coffee.
The break room smelled like a coffee maker that had run all day: old grounds, warm plastic, the faint sourness of burned sediment. The kombucha tap was copper-trimmed and entirely unused at this hour, standing there like a feature in a pitch deck. Priya poured a cup she didn't particularly want and leaned against the counter.
She went through the standard explanations because that was the job. Training data contamination -- a batch of mislabeled data could produce systematic miscalibration. But contamination produced drift at the category level, not seven precise high-confidence selections of specific image details. Contamination scattered. It didn't pick a dog's eyes and a sidewalk crack with 97.1% and 94.8% certainty. Adversarial inputs didn't fit either -- the seven images were standard street-level photography, nothing engineered. Edge detection artifacts produced noise at low confidence, not 94.8%.
She drank the coffee. None of the standard failure modes fit. She had been trained to work through the list before considering anything outside the list, and she had worked through the list, and the list had come up empty. What fit the data was the explanation she was not going to write in the ticket.
What fit the data was that SELECTOR-9 was selecting things that interested it.
She had taken the GridTrust training module on anthropomorphizing AI models eighteen months ago, the same one everyone in QA took. She remembered the main point: humans were pattern-recognition systems, and one of the patterns humans were most prone to recognizing was intentionality. We saw faces in clouds. We attributed motives to weather. When a model produced an unexpected result, the human tendency was to ask what the model wanted rather than what the model's architecture had produced under specific conditions. Anthropomorphizing was the wrong question. Behavior had technical explanations. The job was to find the technical explanation. She rinsed her cup and went back to her desk.
By morning, Conference Room B smelled of old coffee and the dry-erase markers no one capped properly, and she had the slides up on the wall display before Nate arrived: seven images arranged in a grid, the grid coordinates and confidence scores labeled in the standard font. She had added a second slide with the latency distribution and a third with the confidence contradiction matrix -- selection confidence plotted against object-match confidence for all 12,000 images, the seven anomalies sitting in the upper-left quadrant where nothing should be.
Nate came in with his own coffee, closed the door, and sat across from her. Complete attention, no visible reaction until he'd processed what was on the screen. "Walk me through it," he said.
She did. Twelve thousand images in the audit window. The latency flag, already on the ticket. But she'd noticed the pattern in the processing distribution, pulled the full selection history, written the query. Seven results. She went through each image: what was in the selected square, what SELECTOR-9's confidence scores showed. Nate asked questions as she talked -- good questions, the kind that indicated he was actually tracking the data. He asked about query logic, whether she'd ruled out a logging error, whether the seven images shared anything in their metadata that might point to a data pipeline issue. She had checked all of that. She walked him through it. When she finished, Nate looked at the confidence contradiction matrix for a moment.
"Seven anomalous selections in twelve thousand images," he said. "That's 0.058%."
"Yes. But the confidence scores--"
"High confidence, wrong object. I see that." He looked at the images again. The woman, the balloon, the dog. "What are you thinking the explanation is?" It was a genuine question, not a trap.
"I don't know how to characterize it technically," she said. "The standard failure modes don't fit. Training contamination doesn't produce this distribution. Adversarial inputs don't look like this. The confidence is too high for edge artifacts."
"Then what does it look like to you?" She could feel the training module sitting somewhere in the middle of that question.
"The selections cluster around things that have--" She stopped. Tried again. "Things with movement, or contact, or the sense of being watched. The dog is looking at the camera. The curtains are outside the window they belong in. I can't find the technical category."
Nate nodded -- he'd heard the thing she hadn't said, and he named it for her, not unkindly. "You're anthropomorphizing. We've all done it -- it's a documented bias, which is why we have the training module. You see a pattern in the selections and the pattern looks like preference, so you assume the model has preferences." He set his coffee down. "Seven in twelve thousand is still noise, Priya. The confidence scores could reflect a local calibration drift that happens to look patterned. It doesn't mean the model is making aesthetic choices." He said the last two words carefully -- not to mock her, but to put them on the table so she could hear them clearly. "Schedule the recalibration. That's the right call here."
She had no better framework. She had the data and she had the pattern and she had the thing she couldn't say, and Nate had math on his side, and the math was correct. She said okay.
She submitted the recalibration request at 11:14 a.m. The workflow was straightforward: anomaly ticket number, model designation, anomaly type, priority level, estimated timeline for the training team. Five business days. She filled in each field. The system accepted the request and sent an automated confirmation. The ticket moved from Open to In Progress, and she sat at her workstation for a moment without touching anything.
Nate was not wrong. Seven out of twelve thousand was a number she would have dismissed herself, looking at someone else's audit. The confidence scores were anomalous but not inexplicable; local calibration drift was a real phenomenon and she'd documented three cases of it in the past two years. She had followed procedure. She had done the job correctly.
She opened the image files again. The seven of them at the bottom of her screen: the woman, the balloon, the pigeon, the crack, the hand, the dog, the curtains. She looked at them for a moment the way she'd looked at them the night before, working for the technical category that connected them. Then she created a folder on her personal drive and dragged the images in. She copied the confidence matrix, the raw query output, her analysis notes. She closed the folder. She hadn't decided to do it; her hands had done it while she was still looking at the images. She closed the analysis windows and pulled up the next item in the audit queue.
The recalibration was scheduled. The anomaly was noted as resolved. Everything was proceeding correctly.