It wasn’t reading enemy screens. It wasn’t injecting DLLs. It was a —a machine that learned the “grammar” of a VALORANT match so perfectly that it could forecast the future five seconds ahead. A stochastic parrot of the server’s own logic.

For six months, XG had been the nightmare of the Pacific League. Undefeated. Forty-two maps straight. Their IGL, “Zen,” called rotates before the enemy even planted. Their duelist, “Raze,” flicked to heads that were still behind smoke. Analysts called it “intuition.” Pros called it bullshit. But no anti-cheat—not Vanguard, not even the invasive kernel-level stuff at Masters—had ever flagged a single XG player.

In the final map of the series—Split, XG’s best map—it happened. Zen called for a B execute on a standard pistol round. The predictor said “two in heaven, one back site.” Raze swung.

No one was there. Three Furia players had stack-planted A, a textbook anti-prediction. XG lost the round. Then the half. Then the match.

The program didn’t look like a cheat. It looked like a neural network overlay—a translucent web of nodes that mapped the server’s tick architecture. Within seconds, it had scraped the past 100 rounds of a random ranked match. Then it did something impossible: it simulated the next 100 rounds, predicting every peek, every utility line-up, every death, with 98.7% accuracy.

Kai extracted the zip to an air-gapped machine. Inside: one executable, no documentation. The file’s metadata was a single string: “XG VALORANT UNDEAD – because you can’t kill what sees the future.”

Lethe was a feedback loop. Every time XG used the predictor, the model ingested that round’s real outcome and updated itself. It grew sharper. But it also left a quantum signature in the server logs—a mismatch between input latency and reaction time. A ghost in the machine. Riot’s anti-cheat couldn’t see the program, but it could see the statistical anomaly: a team whose average reaction time was 80ms faster than human peak, but only on rounds they won .