Miners embed questions through a frozen Qwen3-0.6B backbone and route via a trainable linear head (~10K–130KB .npz), selecting models based on question content and price. Routing calls execute inside a metering proxy within an Intel TDX confidential VM, producing hardware-attested BenchmarkProof artifacts that include question IDs, routing decisions, results, costs, and a TDX DCAP quote. Validators collect proofs at a deterministic block, verify DCAP signatures, measurement registers against approved runtime images, content hash bindings, head commitments, answer grading, exploration integrity, and cost consistency.
Validators pool exploration results by coldkey to build a reference frame (upper convex hull of model accuracy vs. cost), score each miner as accuracy headroom above the frontier, dedup near-duplicate heads, apply evidence accumulation and burn-in, compute weights via Yuma consensus, and publish full-epoch reveals (results matrix, routing decisions, scores, frame samples). Miners download reveals, retrain heads, and commit improved versions on-chain for the next epoch.
