The task
Preparing the guided trace
Loading the sanitized research sample…
The trace could not be loaded
Try reloading this page.
Annotated research trace · one canary run
How a model searched for a proof of one olympiad inequality
·
2reasoning rounds
30proposed next steps
12investigations
8full answer attempts
0grader results logged
How the run works
From one paragraph to a beam of proofs
Each round repeats the same five-stage loop. The beam creates breadth; it does not rank mathematical truth.
Unverified proof attempt
Displayed final
At least one required proof step is only asserted
Branch-by-branch
Follow the exploration
See the visible reasoning, every proposed approach, the sampled investigations, and the answer drafts produced in each round.
Displayed
Sampled
Not sampled
Control
Important: sampled means randomly retained by the configured policy. “Not sampled” does not mean wrong, and “displayed” does not mean best.
Eight generated attempts
Compare the answer attempts
Each card is a complete answer generated from a sampled branch. Branch labels describe the steering instruction; an answer may ultimately use another route.
About token counts: generation totals include omitted hidden reasoning, so they are much larger than the visible answer text. They measure compute volume, not answer quality.
Methods & provenance
Run details
High-level settings and diagnostics that help interpret the trace, without publishing internal infrastructure or hidden reasoning.
Identity
Sample overview
Search shape
Protocol, in plain terms
Recovered diagnostic
Included
What this public view contains
Excluded
What was deliberately removed
Supplied evaluation guide