PRISMPRISM
Scoring, rewards, and weights
How PRISM converts architecture and training evaluation into miner rewards and Platform weights.
prismscoringrewardsweights
Scoring components
Scoring should separate architecture merit from training execution quality and operational safety.
| Component | Use |
|---|---|
| q_arch | Architecture quality after deduplication, review, and benchmark behavior. |
| q_recipe | Training recipe quality for a selected architecture. |
| Benchmark performance | Heldout tasks, long context, reasoning, and robustness. |
| Efficiency | Parameter, memory, latency, and GPU-hour efficiency. |
| Reproducibility/safety | Build consistency and policy compliance. |
Weights contract
PRISM returns positive hotkey weights through `/internal/v1/get_weights`; Platform normalizes them with other active challenge emissions.
json
{
"weights": {
"5F...minerHotkeyA": 0.72,
"5G...minerHotkeyB": 0.28
}
}