SN81 Relaunches with Market-Based Rollout Pipeline

SN81 has been rebuilt as Reliquary with a restructured incentive model that converts the rollout pipeline into a competitive market. Instead of miners generating unlimited rollouts, only the first 16 valid submissions per window enter the batch, forcing miners to predict which prompts sit at the policy's learning frontier before committing compute. This ex-ante filtering approach, inspired by ByteDance's DAPO research, targets 2–3x training efficiency gains over reactive baselines by making gradient discovery a core optimization problem rather than raw generation volume.

• •Rollout batch capped at 16 FIFO submissions per window with sigma≥0.43 zone filter
• •Late or out-of-zone submissions displaced; miners absorb generation cost, not validators
• •Ex-ante frontier prediction required; miners compete on prompt selection accuracy
• •Validator live on UID 237; targets GRPO training efficiency parity with DAPO