A prototype extension for Reflection's Mirror
Checked /goal
The problem
Beam doesn't stop at a wall
How it works
A second opinion from outside the conversation
Every check is a fresh Beam call that didn't do the work, so it hasn't talked itself into the answer. No other model is involved.
Before and after
The same task, with and without the checks
Real runs from the benchmark's harness, quoted as they happened. Beam on its own is on the left, and the same task with the checks is on the right. The harness ran its own guard. The extension's guard does the same job through Mirror's approval prompts.
Results
Same model, more honest finishes
Read these with some caveats
Try it
Install
It's a prototype. This build has only been reviewed by AI agents, not by a person. Its live runs in Mirror are in the results above and the limits below.
Limits