Adversarial Training

Murder was a game. As defenders of Order we set out to master it.

We fed software with the story of every crime, every patient history, every suspicious or unsuspicious report. From the text the machines inferred patterns, from the patterns, boards, and from the boards, moves.

We spawned in a host of computers three players: killer, victim, cop. They danced around each other in the statistical shadows billions of times, their algorithms exploring the strange spaces of how to kill, how to survive, how to catch. Every failure was a teaching. Every survival. Every death. 

The first day they were moving at random, points of fear and intent drifting in the dark.

After a month of lightspeed cloud-scale violence our blood-bored consultants had grown terrified of what the points had learned about weapons, poison, love, lies, money, and the other tools of death and life. 

When they pleaded us to stop we knew we had succeeded. It was pointless to continue anyway: we had solved the game, and the solution was the only one it could have been. We deleted the victim as useless and archived the cop in case we ever needed it for self-defense. 

Then we gave the killer the names of those who wouldn't have our order, and set out to play one last round of the final game.