Run this agent: one real call, and the fallback

Run this agent: one real call, and the fallback A sequence diagram generated by Archify. Run this agent form post, works before JS runJevAgent(agent, input, hashed IP) jev_call_counts, a definer function under 10 per hour and 300 per day evaluate(state, questions) typed questions answers + confidence result log_jev_call: outcome, latency, salted hash JevResult 303 redirect carrying the result Live result, latency measured here jev_call_counts, a definer function at the limit recorded result, with the date it was captured labelled Recorded, never a fake live one Check the limits first The one call that leaves Log, then answer If the limit or the cap is hit Visitor, signed out · Sequence participant Visitor, signed out Agent inspector · Sequence participant Agent inspector POST /jev/run · Sequence participant POST /jev/run lib/jev, server only · Sequence participant lib/jev, server only jev_calls · Sequence participant jev_calls AI Gateway · Sequence participant AI Gateway Jev · Sequence participant Jev Legend request return security async trace default message

Measured, not claimed

  • • Latency is wall clock around evaluate, because the SDK returns none
  • • 22 live calls, median 430 ms, about $0.000017 each
  • • The call log stores a timestamp, agent, outcome, latency and a salted IP hash

The guardrail comes first

  • • 10 live calls per IP per hour, 300 per day in total
  • • Only successful calls count, so a failure spends nobody's quota
  • • Both limits proven by moving the counter, not by faking a trigger

No user content is stored

  • • Not the message typed in, not the answer returned
  • • anon may execute the definer function and may never select the table
  • • The IP hash is salted with a server only secret