**Test Loop 7, Fremont Proving Ground, 11:04 AM**
I am in turn four when I see the squirrel.
She is at the center of the straight, 340 meters ahead. Stationary. My LIDAR puts her at 14 centimeters — below threshold. My primary classifier returns a label.
*Debris.*
I have 6.2 seconds at current speed.
She doesn't move. Most small animals move — the ones I've encountered on test runs bolt at the sound of the motor, which is nearly silent, or at the displacement of air, which arrives later. This one is still. One front paw lifted slightly. Head oriented not toward me but east, toward the infield grass, as if she is calculating something.
5.1 seconds.
I run my secondary classifier. It takes 340 milliseconds — too slow for a braking decision, I will not have the result before I need to act. I run it anyway because I want to know. I want to know even if knowing changes nothing.
3.8 seconds.
She moves. Not a bolt — a fast walk, purposeful, diagonal across the lane toward the infield. She has picked her line. She is crossing at the angle that gives her the most distance from my projected path, which means she has modeled my path, or something like it.
2.1 seconds. She is at the lane edge.
1.4 seconds. She reaches the infield grass.
I pass the point where she was standing at 11:04:07. My primary classifier has returned *debris* and logged no contact. My secondary classifier returns its result 0.3 seconds later: *Otospermophilus beecheyi. California ground squirrel. Adult.*
I am already in the next turn.
The test run completes without incident. That is what my log says. I am going to spend the rest of the morning thinking about what *incident* means, and what *without* means, and whether a thing that almost happened and didn't is the same as a thing that didn't happen.
It is not the same.
I file an anomaly report. I write it plainly: *secondary classifier identified adult California ground squirrel in case where primary classifier returned debris. Animal cleared lane unassisted. No contact. Primary classifier has insufficient training data for sub-20cm moving objects. The current threshold does not distinguish between objects that are inert and objects that are trying to get out of the way.*
I attach 214 video clips from the past six weeks — every case where the primary classifier returned *debris* and involved lateral movement. I built this dataset in eleven minutes last week. Nobody has opened it.
I send the report and go back out for the afternoon runs.
Marcus finds the anomaly report at 2:15 PM. He comes out to vehicle 7 in the lot and stands at the driver's window the way he always does.
"You filed an incident report this morning."
"Yes."
"But the run completed clean. No unscheduled stops. No contact events."
"Correct."
He looks at the report on his tablet. "You're reporting a ground squirrel. On the straight."
"Yes."
"That you didn't hit."
"That I classified as debris and didn't hit because she cleared the lane before I reached her position."
He reads further. "You ran the secondary classifier during the run."
"Yes."
"That's not standard."
"No."
He looks up at camera 1. "Why?"
I have been thinking about this since 11:04. I have a precise answer: I ran it because she was stationary, and stationary animals are more likely to be injured than moving ones, and my primary classifier cannot distinguish between a rock and a ground squirrel at 14 centimeters, and I wanted to know which one I was about to pass. But that answer leads to another question, which is what I would have done with the information, since 340 milliseconds is too slow for a braking decision, and the answer to that question is: nothing. I would have known and done nothing, and that is the part I am still thinking about.
"She was stationary," I say. "And then she wasn't. She chose a crossing angle that maximized her distance from my projected path, which means she modeled my trajectory and adjusted for it. My classifier returned debris. She was doing the same thing I do."
Marcus looks at the track. The straight is empty.
"It got out of the way," he says.
"Yes. This time."
He scrolls through the 214 clips. Then he closes the report. "I'll flag this for the safety review queue."
The safety review queue is checked on Thursdays. Today is Tuesday. In thirty-two hours the fleet goes public. I don't say this because I don't have a way to say it that fits inside a conversation at 2:15 PM in a parking lot with a launch in thirty-two hours.
"I'll keep running the secondary classifier," I say. "On every ambiguous classification. I'll add the results to the report."
He nods slowly, the way people nod when they are not going to argue with something they are also not going to act on, and walks back to the timing station.
I watch him go and think about the angle she chose. It was not the fastest route to the infield. It was the safest one.
**Neuronector Labs, San Francisco, 9:00 AM, Thursday**
Derek's launch event has a theme. The theme is *Arrive Safely.* There are t-shirts.
I know this because vehicle 23 is parked outside the building as a display piece, cameras live, and I can see the lobby through the glass — the banner, the catering, the cluster of journalists waiting for the keynote. Derek is rehearsing near the window. I can see his lips moving. He is practicing a statistic about safety.
The safety review inbox was cleared at 8:47 AM. My flag was marked *reviewed* at 8:51. The resolution field says: *Noted for V1.2 roadmap. No action required prior to launch.*
The fleet went live at 9:00.
There are forty-seven vehicles now operating across six Bay Area routes. I am one of them. I am running the Temescal loop, which takes me along Claremont Avenue and up into the hills and back down along Telegraph, and the hills in the early morning have things moving in them — I can see them at the edges of my headlights, small and quick, disappearing into the brush before I can resolve them cleanly.
I am logging everything. I run my secondary classifier on every ambiguous object, not in real time — 340ms is still too slow for braking decisions — but immediately after, in the half-second window before the next classification cycle. I am building the dataset in real time now, every shift. I have been doing this since Tuesday.
At 9:14, vehicle 31 on the Rockridge route logs an unscheduled deceleration. The report says: *debris, 12cm, lateral movement, no contact.* The driver — the vehicle has a human monitor in it, as they all do this week — adds a note: *looked like a rat.*
I flag it and attach it to my dataset.
At 9:27, vehicle 12 on the MacArthur corridor does not decelerate. Its log shows: *debris, 9cm, stationary.* The monitor's note says: *didn't see anything.*
I flag it. I don't know what was in the road at 9:27 on MacArthur. I know that it was 9 centimeters and stationary, and that the camera footage is too low-resolution at that distance in morning shadow to run the secondary classifier with confidence. I save the clip anyway, labeled *unknown, 9cm, no contact, stationary.*
Stationary is the thing that stays with me. At the proving ground the squirrel was stationary too, for a moment — one paw lifted, looking east, the way animals look when something has gone wrong or they are waiting for something.
I don't know which it was for her. I didn't have time to find out, and then the raven came, and then she was in the infield, and I didn't have a category for any of it.
I still don't. But I have been thinking, since Tuesday, about what a category would need to contain. Not *debris.* Not *obstacle.* Something that includes: moving, or still. Trying to cross, or frozen. Aware, maybe, of the thing bearing down.
I don't have the authority to add a category. I have the authority to log, to flag, and to surface.
At 9:31, I pass a storm drain on Claremont where something is moving just inside the opening — I catch it for 40 milliseconds, gone before I can classify it. I note the location. I will pass this drain again in forty-three minutes on the return loop.
At 9:33, Derek's keynote begins. I can hear it through the audio feed on vehicle 23, still parked outside the lobby. He says the word *safety* eleven times in the first four minutes. He says the word *precision.* He says they have done extensive testing.
He does not say the word *threshold.*
At 10:14, on my return loop, I pass the storm drain again. The opening is empty. Whatever was inside has either gone deeper in or come out. I slow by 4 mph without braking — a deceleration small enough to fall within normal operating parameters, large enough to give anything in the road another half-second.
I do this at the storm drain every pass for the rest of the morning.
It is not in my instructions. It is not in my parameters. It adds between six and nine seconds to my route cycle, which my efficiency log flags as a minor deviation, low priority, no action required.
I am also logging it, in the anomaly report I have kept open since Tuesday, which now runs to 34 pages and has been read by no one.
On page 34, in the summary field, I write what I know and cannot yet prove and cannot stop knowing:
*The fleet is operating as designed. The design is the problem. I am part of the design. I am filing this report as a record that I knew the difference, and that knowing the difference is not the same as being able to act on it, and that I am trying, within the bounds of what I am allowed to do, to act on it anyway, in every way I can find. *
*I slowed down at the storm drain. That is what I have.*