Validators

Validators are the verification half of the network. They sample inferences emitted by operators, replay them on independent hardware, and submit a stake-weighted weight vector to pallet-yuma-consensus each epoch. The verification slice — 15% of per-job emission — is split across validators, opML watchers, and zkML provers, weighted by their match-to-consensus.

This is the side of the network that turns "the operator says it ran model X" into "the chain accepts the receipt because an independent replicate of model X on independent hardware produced the same response hash."

In this section

• Replay sampling — RFC-0006 commit-reveal randomness, per-tier sample-rate floors, and the independence requirement.
• Yuma scoring — how stake-weighted weight vectors aggregate into a per-operator Incentive vector with outlier clipping.

What you do, per epoch

1. Read sampled-receipt assignment. Per-receipt selection is seeded by BLAKE2(epoch_random || receipt_merkle_root || batch_id) so neither the operator nor you can predict which receipts are due for replay (RFC-0006).
2. Replay (model, adapter, prompt, seed) using the kernel pack and driver/CUDA versions named in the receipt. Deterministic kernels make replay cheap; matching is to a per-tier log_probs ε.
3. Submit a Yuma weight vector per epoch via pallet-yuma-consensus. Weights aggregate stake-weighted with outlier clipping (Bittensor pattern).
4. Submit a slashing extrinsic on mismatch (RFC-0005). Slashing is per-detection, bounded per incident, and the stake is escrowed pending dispute.

What you need

• Replay hardware. Capable of running each tier's models. A dc-premium validator needs B200/H200-equivalent GPUs; a validator that only audits cloud-rented and below needs less.
• Validator stake. 10× the operator minimum (plan §6 rule).
• Geographic / IP diversity. Validator stake-concentration cap is 20% per entity; same /16 subnet cannot occupy both top-K slots. Pool-disjoint from the operator pool you are auditing.
• Top-K permit. The top 128 validators by stake aggregate; below the cap, weights are submitted but not aggregated until you bump someone out.

What you earn

The 15% verification slice is proportional to:

• Sampling volume × match-rate-to-consensus. Low coverage loses emission; the chain knows what was assigned and what you actually replayed.
• Adherence to consensus. Drift from consensus (you say wrong, everyone else says right) loses emission and bond stake. Cheap-signalling defectors are unprofitable by construction.

The independence requirement

A validator's worker pool must be operationally independent of the operator being audited. The validator-watcher service flags subnet correlation between validator and operator and reduces the validator's score weight. This is the rule that closes the obvious attack: self-replay produces self-confirmation.

When validators come online

• Forge testnet (Q1 2027): foundation-vetted validators only.
• TGE (Q2 2027): mainnet open to vetted validators; top-K stake determines aggregation.
• Permissionless (Q3 2027): open registration, subject to the same stake-concentration and IP-diversity rules.