Grail operates as subnet 81 on the Bittensor network, serving as the reinforcement learning fine-tuning and post-training layer of the Covenant AI ecosystem. The project implements the GRAIL protocol (Guaranteed Rollout Authenticity via Inference Ledger) to provide cryptographically verifiable model outputs during reinforcement learning rollouts. Grail transforms foundation models from pre-training subnets into deployable, aligned AI systems through permissionless, incentivized RL fine-tuning infrastructure. The platform generates thousands of rollouts per second using custom algorithms designed for post-training optimization. The subnet supports models ranging from 8B to 70B+ parameters and integrates with multiple RL environments for diverse training scenarios.
