The learning-frontier market, opened, then generalized.

1. nowphase 1
   now · the learning-frontier marketlive now

   Miners find the prompts the model learns from

   The market is live on Subnet 81. Miners compete to select prompts at the policy's learning frontier — group-σ in the trainable band, not yet in cooldown. The validator verifies every rollout with GRAIL, runs a GRPO step on what survives, and publishes the updated checkpoint to Hugging Face every ten trained windows. The live policy is Qwen3.5-4B across mixed OpenMath + OpenCode environments.

   • ▲Ex-ante prompt market — miners bet their own compute on the σ-trainable band, replacing DAPO's reactive generate-then-discard filter
   • ▲GRAIL anti-fabrication proof on every rollout — the validator recomputes the forward pass and pays fabricated work zero
   • ▲Live GRPO loop — PPO-clipped, KL-penalized; checkpoint published to HF (ReliquaryForge/qwen3.5-4b-reliquary) every ten windows
   • ▲Controlled result — pass@1 0.470 → 0.610 (+14pp), about 2× the training efficiency of vanilla GRPO

   shipped result · held-out pass@1
2. nextphase 2
   next · open the marketdesign in flight

   Bring your model and environments

   Open the market to external workloads. A client brings a model and a set of environments; the miner network handles rollout generation, frontier selection, and verification; the client receives an optimized GRPO training signal — verified, advantage-weighted, no infra to run.

   • ○Client-supplied model + environments — point a training run at the network and get verified rollouts back
   • ○Sandboxed reward functions — deterministic exec, resource caps, no network
   • ○Per-job isolation — external workloads never collide with the canonical training run
   • ○GRAIL proofs returned per rollout — re-verify offline before feeding your trainer
   • ○Pricing — per-accepted-rollout TAO with credit packs for sustained workloads
3. laterphase 3
   later · a general-purpose RL inference layeropen design

   Decentralized RLHF as shared infrastructure

   Generalize Reliquary into a general-purpose RL inference layer on Bittensor: many teams point training runs at the network at once and get verified, frontier-selected rollouts at scale. The market becomes shared infrastructure for decentralized RLHF — selection intelligence priced as a first-class commodity.

   • ○Multi-tenant trainers — many concurrent training runs sharing one miner network
   • ○Open environment registry — community-contributed environments and graders
   • ○Cross-run miner reputation — selection skill that transfers across workloads
   • ○Economics — emission split across the canonical training run and external jobs

Find the frontier · prove the rollout · train the checkpoint.