The ruling moves the fee. Plenty of marketplaces will show you an AI verdict on a dispute. Almost none let it touch the funds, because the moment it does, every weakness in it costs somebody real money. A verdict that changes nothing is a decoration.

One field goes to consensus. RELEASE or REFUND. The amounts are never asked: the contract computes them, because asking a model to divide a fee is asking witnesses to agree on arithmetic the contract can do exactly and for free.

Nobody's money is locked without a way out. If the worker never delivers, the client reclaims the fee after the delivery deadline. If the client never answers a delivery, the worker claims it after the review window. Both are enforced by the chain rather than by anybody's agreement, so a party who walks away cannot freeze somebody else's money.

The deliverable is pinned. A delivery can carry a URI and a digest computed off chain, recorded in the same transaction. The contract cannot fetch it, and a round that tried would time out on this validator set, so it does the one thing it can: it fixes what was handed in, so the thing being argued about cannot be swapped afterwards.

An unreadable round settles nothing. Not a default and not a coin flip. The job stays disputed and can be put to the network again. A default would be a decision nobody made, taken with somebody else's money.

The complaint is not evidence. The prompt says so: it is the client's account, and if it describes a fault that is not actually in the delivery, that alone is reason to release. Otherwise a client could win by complaining loudly.

This runs on Studionet, deliberately. On testnet-asimov emit_transfer records a payout and leaves balances untouched, measured across two deployments. A contract whose whole point is that money moves cannot demonstrate it there.