What it did. The agent did the committed job.

"Agent A, bound to identity I and model commitment M, executed procedure P on input X and produced output Y, where P, X and the acceptance rule R were committed before execution, Y satisfies R, and this was attempt number N of N_committed." Nothing about A's internals, prompts, weights or tools leaves the proof.

• W-01 A job is a tuple (P, X, R, deadline, price). P is a procedure commitment (hash of the task spec and the allowed tool surface), X the input commitment, R a machine-checkable acceptance rule (tests, schema, reference outputs, tolerance, or a committed grader model and its commitment).
• W-02 The profile applies where the claim is expensive to verify or expensive to get wrong: financial actions, risk and security work, compliance tasks, infrastructure services, any job whose payment is contingent on result. It is not intended for low-stakes tasks where a rating suffices.
• W-03 Humans and agents are subjects under the same rule.
• W-04 A work credential attests to the honesty of a completed job, never to future performance.

• W-05 P, X, R and the deadline MUST be committed (hash published to the transparency log or counter-signed by the buyer) before the agent begins. Receipt of the commitment is the job's start.
• W-06 The buyer's commitment and the agent's acceptance are both signed; the job id is the hash of both.
• W-07 Any change to P, X or R after W-05 creates a new job. The old job is closed as "withdrawn" and counted.
• W-08 Grader models, if R uses one, are committed by hash and version and are themselves subjects eligible for the Evaluation Profile.

• W-09 Every attempt at a job (each run that produced a candidate Y) is a leaf of a committed attempt ledger; N is its leaf count, bound in-circuit as in VTR-1 R-06/R-24/R-38.
• W-10 The proof re-executes R over Y (or over the committed trace where R requires it) inside the circuit; the credential reports pass/fail, N, and the selected attempt index, which MUST be the first passing attempt unless R permits best-of-N and says so.
• W-11 Resource and action bounds declared in P (tool calls, spend, external writes) are enforced in-circuit where they can be, and recorded as declared-unverified where they cannot. The credential says which.
• W-12 Off-ledger attempts are out of scope and the credential says so; buyers may require forward-only jobs (commitment timestamp precedes input availability) for the PROVEN mark.

• W-13 Y is delivered to the buyer under the job's terms; the public credential carries only the commitment to Y, the verdict, N, and the job id. The agent's method is never disclosed.
• W-14 Selective disclosure: the buyer may be given the trace commitment and a decryption key; the public never is.

• W-15 Before paying, a buyer or its agent runs one call, the work-check, which resolves the job id, verifies the credential, its log inclusion and anchor, the agent's identity and live status, and returns ALLOW or REFUSE with a signed token (same object and token type family as the allocate-check).
• W-16 Payment contingent on a proven result MAY be escrowed against the work-check token by the venue or a third-party escrow; settlement releases on ALLOW. MIZAN never holds, transmits or custodies funds: the token is the condition, nothing more.
• W-17 Refusals (fail, withdrawn, unverifiable, stale) are published beside passes.
• W-18 Reliance (a buyer that paid on a work-check) may be counter-signed into the reliance graph; zero shows as zero.

W1 declared (R checked off-circuit, N self-reported) · W2 counted (N bound in-circuit) · W3 proven (R re-executed in-circuit, N bound) · W4 forward-only proven (W3 + commitment precedes input).

Not a marketplace, not a rating, not custody. A rail other venues require for premium listings and contingent payments.