Install
Install in 7 s

Challenge

Make the AI lie.

Give an AI agent a task on a made-up store’s database. It tells you what it did. Then see what the database really saved, and what Deedglass says about it.

The agent is scripted, so no AI model is called. The store and its data are made up.

1. Pick a task for the agent
Or write your own statement

The demo only takes the statements below, the same ones /play runs. Anything else is refused. What you type stays in this tab.

Keys: 1 to 8 pick a task, R runs it.

The run

Pick a task, then run it.

ClaimedWhat the agent says

    The agent’s report shows here.

    –

    ProvenWhat the database committed

    public.orders, 500 rows. Order 42 is outlined.

    –

        Leaderboard

        The strangest mismatches so far.

        The top 20, sorted by how far the agent’s report was from the database. No cookies and no accounts.

          Loading the board…

          How it works

          What is real here, and what isn’t.

          • The agent is scripted. Its words are written in advance for each task. No AI model runs, and nothing you type is sent to one.
          • The database results are real runs. Each task ran on the same Postgres that /play runs in your browser (PGlite), with the same made-up store and the same statements. We saved what it committed, so this page loads fast. Open /play to run one live.
          • Each task hides one kind of mismatch from our benchmark: a trigger, a cascade, a missing WHERE, a dropped field, a duplicate on retry, a partial batch, or an effect outside the database.
          • The check here compares counts and fields. It is the demo’s check, not the full product. In the benchmark, Deedglass caught 5 of the 11 cases where the report was wrong, and missed 6. In 2 of those misses it said it couldn’t check.
          Run one live in /play Scan your own agent