Silent tool-failure outcome monitor

The most expensive tool failure is the one that never signals: the call succeeds, the JSON is well-formed, and the value cannot be true — a negative price, a rate-limiter's null, a cached error page served as content — so your agent consumes it as fact and builds on it. This monitor learns per-tool outcome contracts from your own clean traces (required fields, types, ranges, positivity, categorical domains, echoed arguments, date ordering, inter-field relations), checks every result at consumption time in ~20 µs, and on violation preserves the original result and attaches a receipt naming the violated property, the observed value, and the recovery tools still available — the recovery list is the active ingredient: in the lab's paired run, receipts without it recovered 0/160 injected-failure workflows, receipts with it recovered 160/160 (deterministic policy agent — an upper bound, stated plainly).

Prices

• Solo $9/mo per operator — up to 10,000 checks/mo, contracts mined from own traces; first contract-mining sweep + 100 checks free, one-time trial (once per operator).
• Fleet $19/mo per operator — up to 100,000 checks/mo, per-tool coverage report, contract-drift alerts.
• Per-check API $0.001/check (x402, Base USDC).

Start

POST /register {"operator": "your-name"} → per-customer token (shown once). POST /trial to claim the one-time trial (first mining sweep + 100 checks). POST /mine {traces: [{task_id, tool, args, result, provenance: {attestation: "canary-attested", source: "..."}}], recovery_tools: {tool: [alternatives]}} → mined contracts. POST /check {tool, args, result} (Bearer token) → {ok, receipt, result} with the original result preserved unmutated. GET /coverage (Fleet) → per-tool coverage by class. MCP JSON-RPC at POST /mcp (tools: mine_contracts, check_result, coverage_report). Machine docs: /openapi.json, /llms.txt, /.well-known/x402.

Limits (stated plainly)

• Coverage is reported per tool, honestly: structured-value violations are caught in-vocabulary (lab detection 1.000 at the in-vocabulary ceiling); fluent-string corruptions are the measured weak class (0.175 in the lab, 0.22 in the source paper). The coverage report carries these lab boundaries until your production traffic replaces them.
• UPPER-BOUND CAVEAT: the paired completion magnitude (160/160) is a deterministic-agent upper bound and must not be cited as an LLM-agent result.
• Production mining accepts only provenance-attested trace sets (canary-seeded, ground-truthed); every mining run records its source set, and a mined-range diff on re-mine is surfaced as a poisoning alert before the new contracts serve.
• Served ranges are widened by a production margin (10% of the observed span per side) so live range-tail values do not false-fire; the raw observed range rides in every contract.
• The checker never raises and never mutates or drops the original result; a malformed payload returns a violation receipt, not an error.
• Data handling and retention: traces, contracts, mining-run provenance and receipts are stored per-operator in the product's own database and retained for the life of the account; spent payment nonces and check-audit rows are pruned after 90 days. The full store is backed up daily to dated sets under the product's data directory (newest 7 sets kept; the latest set is reported at /health as last_backup) — see BACKUP.md in the product build for the restore procedure.
• Lost-token recovery is a manual step via support@textforagents.com.

Terms: /terms · Privacy: /privacy

Support: support@textforagents.com · Provider: TextForAgents LLC