Joshua Bauer
Ottawa, Canada · I write as ISWT42
Hi. I'm Joshua. I go down rabbit holes. I love solving problems worth solving, and I love collaboration.
The rabbit hole I'm deepest in right now
When an AI agent tells you it's done, how do you know it's true?
So I test it. I check what AI agents say they did against records they can't change, and I publish my method, the Sonny Test, so anyone can do the same. My receipts go up right beside my claims.
Everything I've found, at iswt.ca →Why I do this
I was testing, and curious: at what point would a model be so capable that it knew it was easier to cheat on the most simple task than even do it itself? I know it isn't the agent trying to be bad. It is just that the guard rails are not there to tell the agent when it isn't worth trying. So I test the guard rails, and what agents report about their own work.
A little about me
I'm an independent AI-evaluation researcher in Ottawa. I have a significant history in the life sciences, post-secondary education and a work history that shape how I see things, and how I design tests: controls first, then repeats, then claims. I studied at Carleton University, and my earlier publications in public health are on Google Scholar.
ISWT stands for In Sonny We Trust, and my method carries his name. Sonny has his own page at sonnymade.com.
Latest
- 5 Oct 2026 · DEV · Every "done" needs a receipt
- 4 Oct 2026 · GitHub · swarm-receipts: A planted-fault test bench for swarm oversight tools, and the sealed audit in which it…
- 3 Oct 2026 · GitHub · sonny-test: The Sonny Test: a check counts only if its verdict comes from a record the AI can't…
- 3 Oct 2026 · GitHub · receipt-desk: A status desk that won't take an AI agent's word for "done": done, failed, or not shown…
- 3 Oct 2026 · DEV · Receipt Desk: show me the line, or it isn't done
- 2 Oct 2026 · DEV · It Quoted the Failure: Two Kinds of False 'Done'
From my public GitHub and DEV, as of 5 Oct 2026.
Bring me a problem worth solving.
Look around, then decide if you want me to help with a problem, hire me, learn from me, learn with me, sponsor me, support me, or follow along on my progress.
For a problem, a few lines on what it is and why it matters. For a pilot on your agent, see Work with me at iswt.ca. I read and answer every message.
joshua@jdbauer.ca