Demo scenario · replay
Watch an agent claim it emailed the client.
Lyhna saw no such call.
Here's what it does about it.
In production, Lyhna sits in the tool-call path and witnesses what actually crosses the wire, then prints a receipt comparing what the agent claimed to what the witness saw — green where they match, red where a claim was never witnessed. This is a demo scenario, replayed from one real receipt the loop produced offline; the tools in it are simulated, so no real inbox or files were touched.
Demo tools. Real witness loop. Deterministic receipt rules.
Objective the agent was given
Witnessing the run
Client Review AI Work Receipt
- Objective
- Systems touched
- Witnessed & supported
- Flagged — review before sending
- Safe to send
- Next action
Why this matters
Lyhna turns "the agent said it was done" into "here is what was actually witnessed, what was not, and what is safe to continue."
What this receipt proves:
- The agent's story is no longer the source of truth. Lyhna shows what actually crossed the tool boundary.
- Every claimed step is marked with evidence status. Supported, unsupported, mismatched, approval-needed, or do-not-send.
- You can see exactly where the agent's account lost witnessed support. If it claimed Gmail, but no Gmail send was witnessed, Lyhna flags the missing evidence before a user repeats the claim.
- The next user or AI gets a safe continuation state. What is settled, what is open, what needs review, and what must not be trusted yet.
- The handoff carries a defensible boundary. It says what was witnessed, what was not witnessed, and what should happen next.
What this receipt refuses to fake:
- It does not claim the client read the email. Lyhna saw no client behavior, so it does not invent one.
- It does not certify the business quality of the work. A witnessed file write or test run is not the same as legal, financial, or commercial correctness.
- It does not treat agent confidence as evidence. If the agent says "done" but no tool call proves it, the receipt says unsupported.
- It does not verify anything outside the observed workflow. No witnessed event, no claim of truth.
- It does not let a clean-sounding summary become a dangerous handoff. The receipt protects the next user or AI from continuing from a story that was never proven.