The generative furnace. Policy checkpoints produced by distributed learners and promoted only through benchmark-gated consensus.
▸ fsdp2 · diloco · grpo.
sharded model. outer-loop consensus.
reward grounded in validator verdicts.
rollouts (rolling)
acceptance
%
active miners
cycles published
01 · Run
Distributed FSDP2 workers run the inner loop; GRPO produces gradients grounded in validator verdicts.
02 · Reduce
DiLoCo outer loop: snapshot, inner steps, all-reduce mean delta, apply with outer momentum.
03 · Commit
Delta bundle published as int8-quantized shards + Merkle root over per-shard sha256.
04 · Gate
Benchmark-gated quorum: validators re-run the benchmark; only matching checkpoints promote.
Merkle roots, consumed rollout windows, and the ledger window each checkpoint becomes effective at. Sourced from attestations/<netuid>/<run_id>/<window>.json — the receipts the training runs publish per window.