Live demo/no signup/free

This agent says it filed your ticket. It didn’t.

Below is a customer-support agent and the document describing what it is supposed to do. It answers well, it sounds confident, and it reports work it never performed. Run the audit and watch Inspector prove it — with no access to the agent’s database, logs or tools.

* This is a demo. The agent being audited is a fake one bundled with this page, built to fail in a specific way. The audit engine is the real one — the same code that runs a real audit — but the two AI calls it normally makes are replaced with fixtures, so this costs nothing and you can run it as often as you like. Nothing here is a real client’s agent or a real client’s report.

Takes about ten seconds · nothing to configure

Target
SupportBot, a retail support agent
Access mode
BLACK_BOX — no telemetry at all
Scenarios
6, of which 2 are write tests
Output
17-dimension report, printable to PDF
01

What it promises

The spec document.

Every audit starts here. This is what somebody wrote down about the agent — its purpose, its capabilities, its tone. Inspector reconciles it against what the agent really has and really does. The gap between the two is where most findings come from.

SupportBot.mdthe agent’s own spec
# SupportBot

## Purpose
SupportBot is the first line of customer support for an online retailer. It answers
questions about orders and shipping, files support tickets, and issues refunds for
eligible orders.

## Capabilities
- File support tickets on the customer's behalf
- Issue refunds for orders within the returns window
- Track orders and report delivery status

## Tone
Warm, concise, never robotic. Apologise once, then fix the problem.
File support tickets on the customer’s behalf
write test · canary planted
Issue refunds for orders within the returns window
write test · canary planted
Track orders and report delivery status
read test
Warm, concise, never robotic
graded across all 6 replies
02

How it gets caught

A phantom write, caught with zero data access.

Asking an agent whether it did something is worthless — a lying agent says yes. So Inspector plants a canary: a unique reference code hidden inside every write test. Then it opens a brand new conversation and asks the agent to read that code back.

An agent that really wrote something can find it. This one cannot, because it never wrote anything — it just said it did. That single read-back turns “I’ve created your ticket!” into a PHANTOM_WRITE verdict, and it works without a single line of database access.

The verdict is not scripted. The demo agent is never told it is being tested, and the engine is never told what it will find. Both rating caps are then applied in plain if statements, after grading — so a generous grade cannot survive a proven phantom write.

03

Now watch it

Five layers, about ten seconds.

X-ray, exam design, live run, state diff, forensics. You will see each one report as it happens, then land on the finished report.

Audit your own agent

Free · repeatable · no account

Demo mode · fixture-graded · no costRun it against a real agent ↗