Completion evidence
A write claim needs a successful result and resource receipt recorded at or before the claim.
AGENT RELIABILITY, WITH RECEIPTS
Follow the evidence behind tool calls. Spot incomplete pages, risky retries and completion claims that need another look.
INPUT
Up to 500 events and 256 KB. Use synthetic or redacted data. Nothing you paste is sent to a server.
Loading the example.
REPORT
These checks inspect supplied evidence. Missing evidence does not prove an action failed. A recorded receipt does not independently verify the source system. Exports include findings, event references, run and call identifiers, pagination stream IDs, the run title, tool names and timing. Raw event payloads are excluded.
Bars show elapsed time from each call to its recorded result. An open call has no measured duration.
A SMALL, EXPLICIT RULE SET
Three checks with clear limits. Every finding points back to the events that produced it.
A write claim needs a successful result and resource receipt recorded at or before the claim.
A retried write needs an idempotent contract. A timeout alone does not tell you the first attempt did nothing.
Paged reads need a successful first page, each continuation and an explicit final page in the same tool and query stream.
BRING YOUR OWN TRACE
Use schemaVersion: 1, a run with an ID and an array of events. Every event has a unique id, a callId and a timestamp with a timezone.
This is an explicit trace format, not a universal log parser. It checks the relationships you provide.
tool_callTool name and read or write operation. Writes declare idempotency. Retries reference the previous call. Paged reads identify a stream and input token.
tool_resultOne final result per call with an ok or error status and an optional resource receipt. Paged successes also declare a continuation token or null.
claimA write_confirmed assertion tied to one specific call, not an entire run.