A change passes its check. A reviewer asks a plain question of it: can an upload read a private file on the server?
What does the passing check actually tell you about the answer?
In four turns, follow one recorded review: three pull requests that each said the check was now sound, and the reviews that each found it was not. The author claims; the reviewer edits the recorded evidence and runs the author’s own checker again. Each finding stays hidden until you accept the question, and each one opens onto the evidence: the record the reviewer changed, the review’s own words, and the checker’s result.
This research prototype explores how to make agent decisions reviewable: connect a claim to a rule, a result to evidence, and an exception to the person and proposal it covers.
How to follow the review: choose Solve to ask the reviewer, Tab to accept a suggested question, then the → arrow to continue. Each part is a session; each lesson is a turn.
After the review, choose the question you want to investigate:
| Your question | Where to look |
|---|---|
| Can a check tell a real file read from a forged one? | The first review |
| How many reviews does one honest check take? | The last commit, which nobody reviewed |
| Is a shared image what it says, before another application opens it? | Images, Proton Drive and the Rails canary |
For a team assessing an agent workflow, the practical questions are the same: what must hold, what observations would reveal a violation, and who can authorize an exception?
Next Steps & Collaboration
- Run the demo: Start the review
- Inspect the evidence: Claim-to-evidence matrix and source repository
- Discuss applying this to an agent system:
- Public collaboration: Open an issue for open discussions and reproducible evaluations (public; describe workflows without private data).
- Private enquiries: Email
public.rant@pm.mefor confidential discussions regarding production agent architectures, consulting, research collaboration, or funding. (Please do not include credentials or unredacted secrets in initial outreach).