Proof of compute (PoC) is how Gonka reaches consensus without staking tokens. Instead of locking capital or burning energy on arbitrary puzzles, participants prove they control real ML compute — and the network verifies that proof cheaply — so the proof itself backs their vote.

Each epoch opens with a competition the whitepaper calls the Sprint. At a block height fixed by the epoch schedule, every participant's ML nodes initialize transformer models seeded by a recent block hash, then search for nonces whose model outputs land within a target distance. Results are batched and submitted on chain as proof-of-compute batch messages (MsgSubmitPocBatch). Seeding the task from a fresh block hash matters: nobody can precompute an advantage, because the task doesn't exist until the block that defines it does.

Because the proof workload uses the same transformer stack that later serves paid inference, a participant can't look strong in the Sprint while being useless at the actual job — capacity claims are backed by the kind of compute that does real work.

After the Sprint, participants switch to auditor mode: each one samples other participants' submitted batches, re-generates the sampled proofs, and posts validation verdicts on chain (MsgSubmitPocValidation). A submission is accepted when more than half of the network's compute weight vouches for it. Accepted results then flow into SetComputeValidators, which sets both consensus voting power and — combined with the epoch-group ledger — how inference work gets allocated for the next epoch.

You can watch proof-of-compute batches and validation messages land in the transaction feed, usually clustered near epoch boundaries. For the full stage-by-stage mechanics, see How Gonka's Proof of Compute Works.