- Pretend to run the model while returning random or low-effort outputs
- Shortcut inference by using smaller models internally
- Replay old outputs instead of generating fresh responses
- Drop computations entirely to save power
- Poor user experience
- Lower model accuracy
- Network instability
- Unfair payouts to dishonest nodes
- It looks like a normal inference request
- It behaves like a normal request
- The node cannot detect that it is a test
- The node must compute it exactly as it would any inference.
- A known prompt
- A known model state
- A deterministic evaluation path
- The system compares the node’s answer with the known solution
- A slight tolerance window is allowed because of quantization variance
- Deviations outside the threshold indicate the node did not compute the request faithfully
- Routing priority
- Eligibility for high-paying traffic
- Participation in Prime Triads
- Share of Npnts-based emission rewards
- The node receives 0 Npnts
- Therefore earns 0 FAR rewards
- Cheating becomes strictly unprofitable.
- The launcher bans it automatically
- The node’s hardware fingerprint is blocked
- The wallet address is added to the global denylist
- Its prime partners (if any) are notified to recover gracefully
Always.
Nodes maximize profits only by performing real work.