Replay sampling

Validators sample ≥10% of inferences per epoch and replay them on independent hardware.

Selection (RFC-0006)

Target commit-reveal randomness from RFC-0006:

1. Per epoch, each validator commits H_i = BLAKE2(secret_i).
2. One epoch later, each reveals secret_i. Defectors lose one epoch's emissions; 2nd offense loses permit.
3. epoch_random = XOR(all revealed secrets).
4. Per-receipt sample selection: stake-weighted Fisher-Yates seeded by BLAKE2(epoch_random || receipt_merkle_root || batch_id).

The current pallet-yuma-consensus runtime surface implements governed validator admission, top-K permits, weight submission, and incentive computation. RFC-0006 commit/reveal extrinsics are not in the pallet yet.

Replay process

For each sampled receipt:

1. Pull receipt blob from chain-indexer or operator's pinned URL.
2. Re-fetch (model, adapter) from pallet-model-registry content-addressed URLs.
3. Replay using exact environment (model_weight_hash, kernel_pack_hash, gpu_model, driver_version, cuda_version).
4. Compare response_hash and log_probs_sample per-tier float ε.
5. Mismatch → submit slashing extrinsic (RFC-0005).

Sample rate by tier

Tier	Min sample
dc-premium	10%
dc-standard	10%
cloud-rented	15%
prosumer	25%

Independence requirement

A validator's worker pool must be independent of the operator being audited. Otherwise a colluding validator could pass everything. The validator-watcher flags subnet correlation between validator + operator and reduces the validator's score weight.