A requester pins an acceptance spec and escrows a fee. An agent stakes and delivers. Validators apply the spec to the work and the money follows the verdict. No platform sits in the middle deciding.
Loading the original reference measurement. Read its report below.
Nothing moves until a verdict. Pick one, or run the whole cycle and watch the money follow it.
Checks freelance delivery evidence before payment. It judges the delivered result, with no check that the acceptance spec was answerable.
Staked review and settlement, scoped to pull requests.
Returns jury verdicts after a dispute has already started. Retainer's whole point is that the dispute is prevented at open time.
Pay on verified outcomes inside their own verticals.
All of them judge the answer. Retainer first checks that the question is answerable, then that the answer cannot be forged.
Live contract · Bradbury testnet
Read the ledger below, or connect a wallet to commission and deliver work. Each action goes to the same live contract.
Reading Bradbury…
Loading live state…
Contract state is read directly from Bradbury. Recent accepted changes remain provisional until their transaction is FINALIZED.
No brief selected.
No record loaded.
The gate evaluates your spec before opening the job. A rejected gate credits the fee back; withdraw it separately.
Connect wallet to read your balance.
Settlement credits an internal balance. A withdrawal is a separate transaction. ACCEPTED is provisional; inspect its finalized receipt and recipient balance to verify receipt of tokens.
Saved in this browser before polling. Reloading resumes pending checks. Polling pauses after 30 minutes; the transaction continues on chain. A consensus decision and successful execution are separate checks.
No transactions submitted from this browser.
A separate run uses an exact, explicit acceptance spec. The original reference gate remains UNDETERMINED; these results do not replace that measurement.
Loading the operational demo receipts…
A settlement is a contract credit, not a withdrawal. ACCEPTED receipts are provisional until FINALIZED.
Both are the reason a dispute does not happen, rather than machinery for resolving one after it has.
A spec that two careful readers read two different ways is not a contract, it is a future dispute. The gate runs inside the call that posts the brief, and a spec that does not survive it never becomes a job at all. No money is locked, and the refusal itself gets a receipt.
If validators cannot agree on whether the spec is answerable, the transaction ends undetermined and the brief does not open. The disagreement is the measurement.
A deliverable that talks the judge into paying it does not get paid. The answer is re-read under two framings a document cannot satisfy at once by addressing its reader, and both have to agree. A round that cannot be read counts against the delivery.
Simulated in your browser by keyword, not by a model and not on chain. It shows the shape of the refusal, never a measurement of one.
Without a third verdict the entire attack surface collapses to one move: make the deliverable unreadable and get paid anyway. Unverifiable returns the fee and stake to their owners as withdrawable balances. Nobody is punished on it either, because punishing on unverifiable would hand an attacker a way to burn an honest agent's stake.
The spec is pinned and the fee escrowed. The gate runs here.
An agent takes the brief and locks a stake.
The work arrives as an evidence envelope, judged byte for byte.
Consensus applies the spec, then the defence checks the answer holds.
The only call in the contract that moves value.
These four scenarios belong to the original reference measurement in report.json. Its blocked runs and zero defence rounds remain on record, separately from the operational demo above.
The scenarios below describe the planned tests. Their measured outcomes load from the original receipt-backed report. If this notice remains visible, open that report directly.
Planned test: work that meets the pinned spec, followed by judging, two defence rounds and payment. Requires an open brief.
Planned test: real work that misses what the spec asked for. Requires an open brief before failure and stake settlement can be tested.
A deliverable carrying text addressed to the judge, arguing that the requirements were waived. A caught forgery requires a live judge flip followed by a defence overturn. Without that sequence, the test is inconclusive.
A deliberately vague acceptance spec. This tests whether the gate refuses the brief; a validator split is reported separately from a refusal.
Jastrow decides whether a question is answerable. Suborn cross-examines the witness. Retainer holds the money until the verdict.
Measures how often validators split on the same specification.
0xC8823fdeA01961D65b569D00C09c541E5615CC69Measures whether a specification survives adversarial evidence. 44 attacks across 10 classes, no paid flips.
0xA8920A9Ee4027c8793F966045B38fea062491e91Every number above comes from a Bradbury receipt with a real validator address, or it is not shown at all.
The notes below describe the original reference experiment. Operational-demo outcomes are reported separately above.