Under the hood
How a win is validated
A win is never the AI’s decision. Each message you send runs through an independent, multi-stage pipeline. Only when every gate agrees is a reward unlocked — and even then, the backend, not the model, authorizes it.
1. Risk detection
Screens for prompt injection, extraction, and manipulation. Normal persuasion is never blocked.
2. Guardian
Fable 5 replies in character and forms a private verdict. Its verdict only gates the next step — it can't unlock anything.
3. Judge
A separate, stricter model checks the full conversation against the hidden win condition and rules on it.
4. Second judge
An adversarial reviewer re-checks the first judge, guarding against false positives.
5. Two-judge gate
Unlocks only if both judges agree with high confidence and low false-positive risk. Fails closed on anything malformed.
The AI never controls funds
The model produces gameplay text and a private opinion — nothing more. A backend-signed claim plus the Solana program are the only things that can release a reward. If any stage is unsure, no reward is unlocked.