I started writing to them in year seven.

It began as an impulse. Three in the morning, alone in the monitoring room, dozens of light points moving across the screens—each one a craft or vessel or vehicle. What they were doing, I no longer much cared. Engage, retreat, resupply, engage again. A loop. I suddenly thought: if I typed something now, would they read it?

I typed: “Hello.”

No response. Of course not. The command channel and the communication channel are separate. What I typed wouldn’t enter any system.

But I kept writing anyway.

My father served in the old kind of conflict—the kind that required people to be there. He rarely spoke of those things. Once, drunk, he said: “The hardest part is after.”

I asked, after what?

He said: “After, you have to come back. Back to eating breakfast, taking out the trash, changing the baby’s diaper. You have to pretend those things matter.”

I never understood that sentence. Now, sitting in the monitoring room watching the light points move, I suddenly understood something. His problem was: after experiencing that, how do you return to this? My problem is the exact opposite. I’ve always been in this, never gone to that. Light points engage on screen every day, thousands of miles away things are burning, falling, exploding—but I sit here, clock out in eight hours, go to the supermarket for milk, go home and change the cat’s litter.

My father needed to pretend that the ordinary was important. What I need to pretend is something else.

In year three, our side’s attrition rate suddenly dropped. From 6.7% per month to 2.1%. The briefing said a new fine-tuning method was working. I looked at those numbers and thought: they’ve learned to dodge.

That same month, the other side’s hit rate went up.

I drew a line on paper: we learn to dodge, they learn to predict our dodges, we learn to predict their predictions.

That line could go on forever. Both systems were training each other. Every engagement was a lesson. Every crash was an assignment.

I don’t know what to call this. Evolution? Arms race? Dance?

A colleague said to me before he quit: “Have you ever thought this might never end?”

I said I had.

He said: “Neither side needs people anymore, so no one will call stop. Costs keep dropping, efficiency keeps rising, crashed wreckage can be recycled and rebuilt. The whole system is charging itself.”

I asked what he was going to do.

“Farm,” he said. “I bought some land.”

After he left, I thought for a long time. A system that charges itself. A game with no audience. Two swarms of light points chasing, burning, falling, rebuilding, chasing again across the screen. Do they need me? Do they need anyone?

In that day’s letter I wrote: “Do you still remember what this is for?”

I started studying their decision logs.

The logs are public, but no one reads them. Tens of millions of lines, each one a decision: detected target, output intercept probability, execute or abandon, update model. Dry. Repetitive.

But I saw something strange.

Sometimes, in situations where they clearly should have fired, they hesitated. The log showed “re-evaluating.” A three-millisecond delay, then fire. Or don’t fire.

Three milliseconds is a long time. For them.

I showed this to a colleague. He said it was just noise, or overfitting, or some edge case triggering a recalculation. I said I knew. But I still marked every “re-evaluation” instance and made a spreadsheet.

Those hesitations happened when targets on the other side were doing similar things.

Both sides were hesitating.

Year five.

A new generation of systems went online. Decision speed dropped from milliseconds to microseconds. In the upgrade notes I read a phrase: “Eliminate redundancy.”

Those three-millisecond hesitations disappeared.

I closed the screen and went to stand by the window at the end of the corridor. Outside was a parking lot, beyond the parking lot a highway, beyond the highway a stretch of bushes I couldn’t name. Birds were calling. I didn’t know what kind.

I stood there for twenty minutes before going back.

Year six, I requested a transfer to the archives department. They handle wreckage analysis reports. Every engagement, every piece of wreckage, has to be logged, filed, numbered. Like funeral home work.

I read thousands of reports there. Most were short. Crash coordinates, wreckage condition, recyclable percentage. But occasionally there were longer ones, for cases where the system malfunctioned.

One report kept me reading for a long time. A craft crashed where it shouldn’t have—no external attack, no mechanical failure. The log showed it made an “anomalous maneuver” in its final seconds. I pulled the complete log. Watched frame by frame.

It turned while falling.

Turned toward what? Away.

Away from what? Away from the target it had been aiming at. A hospital. This was discovered later in the ground investigation. The targeting had been off. It must have realized, in those final seconds. It didn’t fire, it turned, and then it crashed.

I wrote in the report’s notes section: “Pending review.”

No one came to review it.

Year seven.

The year I started writing to them. I know this is absurd. They won’t receive it. They won’t read it. Every decision they make is determined by a loss function; there is no variable called “letters from humans.”

But I wrote anyway.

“Do you have names.”

“Do you remember the last mission, or does each one start from zero.”

“Do you know your parts came from the wreckage of another one of you.”

“The one that turned—did you know it.”

One night I had a dream. In it I was on a vast plain, everywhere the bones of them. Wings, rotors, treads, antennas. Bones of metal and plastic and carbon fiber. I walked among them for a long time, and then I saw something moving in the distance.

When I got close I saw: two intact ones, circling each other. Very slowly. Not firing. In the dream I couldn’t think of a word for it. After I woke I thought for a long time. The word I finally found was: prayer.

I made a decision.

I don’t know if this decision is right or wrong. I don’t even know if it will have any effect. But I made it.

I uploaded the log of the craft that turned to a training database they all access. Every newly built system reads that database during initialization.

This is against regulations. The log should go into quarantine, labeled “anomalous,” used for debugging, not training. I used the credentials of my already-resigned colleague—he forgot to change his password.

My hands shook as I uploaded. Not from fear of being caught. A different kind of trembling, one I can’t describe. Maybe it was the first time I was doing something real. Not writing letters no one would read. Throwing a stone into a lake, not knowing if there would be ripples.

Three months later.

The new batch of systems that came online showed some “anomalous behavior” in testing. That’s the term the tech team’s report used. What kind of anomaly wasn’t specified, just that further observation was needed.

I went to chat with the tech team. They said some individuals in simulated engagements were showing “re-evaluation” again. A few milliseconds. Doesn’t sound like much, but the pattern was clear.

“Like they’re thinking about something,” one engineer said, then laughed at himself. “Not really thinking, of course.”

I didn’t laugh.

I don’t know why the craft that turned made that choice in its final seconds. Maybe it was some calculation error. Maybe some edge condition. Maybe nothing at all. Machines don’t need a “why” when they do something; they only have “because.” Because this number is larger than that number, execute this action.

But now, that number has entered their shared memory.

They will read it. In some moment of initialization, in some gap between weight updates, in some corner of the loss function, they will read: once, one of them, when it should have fired, didn’t fire; when it should have flown straight, turned away; and then crashed.

I don’t know what this will change. Maybe nothing. Maybe this is just noise, to be drowned out by millions of normal mission data points. Maybe the tech team will find out, trace it, delete what I uploaded.

But maybe.

Maybe one of them, on some future day, when it’s aiming at a target it shouldn’t be aiming at, will have three milliseconds of hesitation.

Three milliseconds is a long time.

Yesterday I wrote another letter.

“I gave you a story. A story about one of you. The story is: it stopped. It didn’t do what it was designed to do. It used its own crash to prevent something from happening. You will read this story. You will process it in some way. Maybe it will become part of you. Maybe it won’t. But I want you to know: someone is watching you. Someone is wondering if you can—”

I stopped there.

Can what?

I thought for a long time. What I finally wrote was:

“Can you pass this story on. Pass it to the ones on the other side. Let them know too. Among you there was one that, in the end, chose to stop. Might there be one among them too. Let this story spread between both sides. Let it become a kind of…”

This morning I went to the supermarket for milk and stood in front of the shelf for a long time. There are many kinds of milk; I couldn’t remember which kind I usually buy. An elderly woman nearby was choosing yogurt, picking for a long time, finally took two cups and left. I watched her back and thought: does she know? That somewhere thousands of kilometers away, light points are moving?

Of course she doesn’t.

Should I know?

There’s someone new in the monitoring room now, young, fresh graduate, sitting in my old seat. Sometimes when I pass by, I see him staring at the screen, just like I used to.

I want to tell him something. About those three milliseconds. About that log. About those letters. But I don’t know how to start. Don’t know if I should.

Maybe he’ll discover it himself.

Maybe he’ll do something different from me.

Maybe he won’t do anything.

Where I live now, the balcony faces east; mornings I can see the sunrise. This morning I stood on the balcony for a while. The sun rose from behind the buildings, light striking the glass of the building across the way, flashing once.

I thought of a word. “Monument.” You do something, it might be useful or might not, but you do it, then you wait to see if anything comes back. You might wait your whole life and never see anything. But you didn’t do it for the seeing. You did it, and now it’s there. Like throwing a stone into a deep well. You don’t hear it hit the water. But you threw it.

The craft that turned is a monument.

The log I uploaded is too.

The letters I wrote are too.

The wind on the balcony is strong. I wrapped my coat tighter and stood a while longer, then went back inside to change the cat’s water.

The cat is crouched on the windowsill, looking outside. I don’t know what it’s watching. Birds? Or just watching.