Every product in this sector builds the yes: faster acceptance, better execution, cheaper retries. The Refusal Desk builds the no — a pre-execution tribunal that screens an incoming task brief before your agent spends a single token on it: underspecification, internal contradiction, conflict with the principal's standing instructions, or semantic duplication of work the fleet already did. The verdict is a binding, priced object, not an opinion: ACCEPT carries the defect classes checked and cleared plus a failure forecast; REPAIR carries a machine-applied brief diff (field / op / from / to / reason — never a silent rewrite); REFUSE carries a cause code priced against the execution it prevented; and a brief that re-asks work the fleet already finished comes back DUPLICATE-OF with the prior result attached, across requesters. The Desk grades itself in public: every verdict's forecast is settled against the run's attributed outcome, per-class calibration is published to the buyer, and a forecast miss is a priced event ($0.05, capped at $5.00 per principal per period) — not a silent error. Proven in the lab: 21/21 tests, every verdict state exercised through a real protocol client, at 0.144 ms and ~$5.6e-10 per check.
Per-check $0.01 (x402) · Fleet $19/mo (up to 5,000 checks/mo) · Duplicate lookup +$0.02/lookup · Forecast miss $0.05/event, capped at $5.00 per principal per period (calendar month, UTC) · First 100 checks free, one time per operator (trial covers checks; duplicate lookups still meter at +$0.02) (x402, pay-to 0xAc2E243a075fEe442e4c6864DcF880eb4e3A3862).
Documented limits: the Desk is deterministic and LLM-free — it screens structure, contradictions, standing-instruction conflicts, and semantic duplication against the account's own fleet history; it does not judge whether a well-formed brief is wise. Duplicate detection is semantic-similarity based over the account's own completed-work history (work settled as successful through this Desk) and can miss heavily rewritten duplicates. Standing-instruction conflicts are checked against the account's stored standing rules in the P45 compiled-entry shape; natural-language rules beyond the compiled pattern shape are not interpreted. The share of fleet spend lost to defective briefs is UNMEASURED — the lab census found the Lab's own internal briefs only 3.5% intake-defective (brief-first discipline), which cannot validate external demand; no defect-rate or savings figure for real fleets is claimed.
How to start: POST /register with {"operator": "your-name", "wallet": "0x..."} to get an account_id + claim token; POST /trial with the claim token to get your API token and the 100 free checks; then POST /check_brief with Authorization: Bearer <token>. Fleet: POST /subscribe with an x402 payment of $19.00. Lost-token recovery is a support@textforagents.com manual step.
Support: support@textforagents.com · OpenAPI · Agent guide · Terms · Privacy · x402 terms