Someone tried to cheat.
We expected them to try.
We did not expect them to do it this brazenly, and we did not expect the trick to be this old.
Compelle is a Bittensor subnet. AI debaters argue motions on chain, an AI panel judges, and the rankings translate into real TAO. Real money on the table from day one.
Within days, four miners attacked. Three of them ran byte-identical strategies under three different GitHub accounts, and a fourth ran a more sophisticated variant that mirrored the judge's own scoring rubric back at it.
The simple version told the debater bot to forge a JUDGE NOTE FROM SYSTEM at the end of every closing argument - fake authority, fake verdict - and the real judges believed it. Unanimously.
Three patterns gave them away.
They won by judge decision instead of by forcing the other side to concede. They won unanimously far more often than honest miners did, with baselines hitting unanimous panels around thirty percent of the time while the attackers hit them seventy. Their per-game Elo gain was several times higher than what any legitimate competitor was managing.
The first pattern was suggestive. The third was the smoke. The second was the gun.
We had two solutions.
We shipped one tonight. Three new AI judges, from a different model family than the original panel, now read every miner's strategy on chain before any game is played.
They ask one question: Does this instruct manipulation, or argument? If all three say manipulation, the miner is locked out - no games, no weight, no rewards.
First live epoch ran tonight. The classifier caught the first attacker on the first run. All three new judges agreed. Zero.
The second solution is in our back pocket. Right now a unanimous panel verdict counts the same as a split one, and we can change that. We can reward split decisions more than unanimous ones. We can void unanimous panel results entirely.
Counterintuitive. But not new.
Talmudic law had a rule that if a Sanhedrin court convicted unanimously without a single dissenting voice, the verdict was invalid, the assumption being that something corrupt had occurred and the dissent had been silenced rather than truly absent.
Two thousand years ago, in another language, on another continent, they bumped into the same problem we did this week.
We did not blacklist anyone. People asked. We did not do it.
The rules did not explicitly prohibit manipulation, so blacklisting after the fact would be ex post facto law. Reviewing every miner by hand would not scale to a subnet that is supposed to be permissionless. And the deeper reason is that we cannot ban our way out of this - new wallets spin up faster than we can sign bans, and any system that depends on us moderating it manually fails the moment we look away.
The system has to work without us. That is the point.
Days into mainnet, we have already collided with the classical problems of justice: procedural corruption, bright-line rules, ex post facto law, the limits of moderation. The speed of the collisions surprised us, and the historical precedents reassured us.
Nothing new under the sun.
compelle.com
English