A dropped subscription looks exactly like a quiet week. This service keeps the per-principal census the protocol draft omits, re-verifies every subscription on a stated cadence, and gives your agent a signed answer - with a freshness bound - to "am I still subscribed, and when was that last proven?"
POST /register (no auth) returns your operator account id and a per-customer token (shown once). Use it as a Bearer token on the MCP tool surface (POST / or POST /mcp, JSON-RPC: initialize, tools/list, tools/call) and on POST /verdict, POST /subscribe, POST /attest-plan. payTo for x402 is the raw address 0xAc2E243a075fEe442e4c6864DcF880eb4e3A3862.
Flatline precision, published as measured in the lab (bench_q2b, exponential arrival law, 600 trials/factor): factor 3.0 (default) false-flags 16.7% of ALIVE streams to UNKNOWN; silent-drop miss 3.7% at 4x mean interval, 0.3% at 5x, 0% at 8x. Silence degrades ALIVE to UNKNOWN (inviting a re-probe), never to a false TERMINATED.
Boundaries, stated plainly: all receipt and observation times are server-stamped - caller-supplied timestamps are never used, so a freshness bound can only be earned by a real probe at a real time; freshness-bound verdicts only; the flatline error table above is the measured one and no other precision figure is stated; the zombie watch exists only on tap shapes (server-sdk / client-proxy) - operator-declared subscriptions get census, probing, flatline and receipts but NO zombie watch and are never described as zombie-watched; real-network probe cost was unmeasured in the lab (localhost floor only) and is measured in production per shape with cadence metering enforced; willingness to pay is unmeasured.
Support: support@textforagents.com