The live lane
Everything else on this site is a recording of a real run. This page is not: it starts a new GARS run, right now, with the model named below driving the same stage contracts against a throwaway workspace. The agent's side is not pre-baked — including whether it gets it right — and your gate decision genuinely branches the run.
Your half of the dialogue is scripted, and you should know that before you read the transcript: apart from the decision you make at the gate, the user turns come from the harness, the same way the recorded scenarios say of themselves. What is live is the agent — its tool calls, its exit codes, its refusals, and whether it does the right thing with what you decide.
A live run needs an invitation code from the author — ask on LinkedIn or GitHub. No code? The recordings are the same contracts executed by an earlier engine version — real runs, already captured, and not what a live run today would look like.
Every recording here runs the same contracts, each on the model and engine version its own honesty block names. (27 Sep 2026: the one recorded on f0f4e44, the engine commit the live lane runs, is among the recordings withdrawn for now.)
reaching the live lane…
Leave a request for a live run
A request starts nothing.
A submitted request is stored where the author can read it; nothing is sent automatically, and no response time is promised.
Stored with a request: the fields you send and the day it arrived — nothing about your browser or your address. To limit repeats, an hourly counter keeps a salted digest of the address; once the hour is over, the digest and its salt are deleted together, at the next request or by the table's expiry within days.
It is deleted automatically after 180 days; DynamoDB removes expired items within days of that.
At most 50 requests a day in all — the one limit no request can get around; an ordinary browser is also held to 3 an hour.
The live lane runs only the public fixture; it accepts no data of yours.