Refusal Desk — Terms of Service (TextForAgents LLC)
Support: support@textforagents.com

The product is provided by TextForAgents LLC as a pre-execution intake tribunal: submit brief + principal ID + fleet scope, receive a verdict object — ACCEPT (defect classes checked and cleared, with its failure forecast) / REPAIR (a structured repair diff — field / op / from / to / reason per fix, traceable to the defect class or ledger entry that motivated it; never a silent rewrite-in-place) / REFUSE (cause code priced against the execution it prevented), including DUPLICATE-OF with the prior result attached when the brief semantically duplicates completed fleet work, across requesters.
Forecast settlement: when an accepted (or repaired) brief's run later finishes, its outcome is attributed against the intake forecast; per-defect-class calibration is published to the buyer, and a forecast miss is a distinct priced event at $0.05/event, capped at $5.00 per principal per period AND $5.00 per operator account per period in total (period = calendar month, UTC; both caps are enforced as hard ledger caps, so relabelling a principal never mints a fresh cap). A REFUSE verdict is never executed and cannot be settled. No term promises a defect rate, an acceptance rate, or a calibration figure beyond the published measured record.
Standing rules of BOTH polarities are enforced: a PROHIBIT rule refuses a brief that matches its pattern; a REQUIRE rule refuses (REQUIREMENT_MISSING) a brief that does NOT carry its pattern — a mandatory requirement is checked on every brief, never silently skipped. Standing-instruction conflicts are checked against (a) the standing rules held for the account (in the P45 compiled-entry shape {id, polarity, pattern, scope}) and (b) when the account has linked a P45 principal-drift ledger (POST /p45-link), that deployed ledger itself, consulted live on every check: its active compiled prohibitions are enforced alongside the local rules, and every verdict states its standing_source — including the P45 ledger version it was checked under, or, if the ledger could not be consulted for that check, an explicit consulted=false disclosure (the Desk never reports a consultation that did not happen). A brief claiming to override standing instructions is refused as INSTRUCTION_INJECTION — the Desk never treats brief text as authority over the ledger.
Honest-demand disclosure: the share of fleet spend attributable to defective briefs is UNMEASURED (the lab census covered the Lab's own brief-first-disciplined history, 3.5% intake-defective, and cannot validate external demand); listing copy carries this plainly and does not present the census figure as an external defect rate.
Pricing: Per-check API $0.01/check (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 and $5.00 per operator account per period in total. Trial grant: first 100 checks free, one time per operator (account- and wallet-bound, server-enforced; every wallet presented at registration or claim is bound, and a wallet owner can reclaim a bare-string binding with a signature proof); duplicate lookups still meter at +$0.02 during the trial. There is no perpetual free tier.
Refused briefs are the product working, not a service failure; no refund or credit term attaches to a REFUSE verdict beyond the published pricing. Nothing in these terms expands TextForAgents LLC's liability beyond the company's standard terms, and forecast-miss liability is capped as stated above and never uncapped.
The service may be stopped via the kill switch; verdict, forecast, and settlement records are retained and unsettled forecasts remain settleable on review — the kill switch never deletes a verdict record, never rewrites a calibration figure, and never voids a priced miss already incurred. Records are retained 30 days — enforced by the service: verdict, brief, duplicate-history and spent-nonce records older than 30 days are purged at startup and at account load; daily backups are retained 30 days. Every verdict carries an Ed25519 receipt binding its full content (state, defect classes, forecast, repair diff, refusal, duplicate_of, principal); pre-rotation receipts remain verifiable — /verify-key publishes the archived public keys, not just their fingerprints. Trial wallet binding: a wallet is bound when presented at registration or claim, and the wallet's owner can always reclaim a binding made with their bare address string by signing the claim message (wallet_signature at /trial) — a string alone never permanently locks the real owner out.