The router at the center of the network.

WorkerManager — Holds every authenticated miner's WebSocket state. Runs heartbeats, computes peer scores, picks the replicas for any given job using a top-half-by-score + random-tail strategy that avoids winner-take-all dynamics.

JobRouter — Fans a single request out to REPLICA_COUNT workers, collects responses under a deadline, and surfaces the verdict to the original caller.

Consensus — Canonicalizes each output as sorted-key JSON, hashes it, and groups the responses. The largest group with size ≥ CONSENSUS_THRESHOLD wins; tiebreaks favor the first response received.

RewardLedger — In-process accumulator. Credits winners on every settled job. Supports snapshot/revert so a failed on-chain submission doesn't lose pending balances.

ChainClient — viem write client. Caches the on-chain batch nonce and submits batchDistributeRewardsevery settlement window.

The orchestrator has only one privileged on-chain action: calling batchDistributeRewards with the credits accumulated in its ledger. It cannot mint $KREW, modify stake bounds, register workers, or pause the contract. The on-chain replay-nonce gate protects against accidental double-submission; the segregated reward pool means even a fully compromised orator wallet can't touch worker collateral.

Today the orchestrator runs as a single Node.js process. Worker state and the reward ledger live in memory. This is intentional for the MVP — it removes a whole class of distributed-systems bugs and keeps the latency floor low.

Multi-process HA is on the roadmap. The replay-nonce check on batchDistributeRewards already prevents two replicas from settling the same batch; what's left is sharing the worker registry and reward ledger across instances (Redis pub/sub is the likely path).