I work on mathematical evaluation, research tooling, and the gap between a measurement and the claim made from it.
My current projects ask practical questions: Does a reward measure the behavior we intended? Can another person inspect the evidence behind an experiment and understand its limits?
Explore research & tools for a guided tour, or start with a repository below.
| Project | What to explore | What is available |
|---|---|---|
| ReviewOps | What changed between evidence snapshots and which version was reviewed | A local prototype with revision comparison, portable review summaries, offline backup/restore, and synthetic cases |
| Math Reward Hacking | How completion limits and answer extraction change an apparent training result | 10,500 saved completions, a worked example, analysis, and verification commands |
To try the workflow tool, download the ReviewOps release. Compare revisions, inspect evidence, and download a readable review summary. Offline backups preserve the complete workspace. It runs locally with Python and no third-party packages.
For a saved-output research example, read the math case study walkthrough. It follows one saved response through different scoring rules and explains why the study does not establish consistent reward hacking.
Formalization evaluation robustness is a documentation scaffold for testing whether equivalent mathematical presentations change an evaluator's verdict. The public branch explains the question and proposed controls; it does not contain a runnable experiment or reported findings.
Each featured repository explains what runs, where the evidence comes from, and what remains unverified. Saved-output reproduction is distinguished from regenerating an experiment; synthetic demonstrations are labeled as such. I use AI-assisted engineering, and treat tests, source inspection, and explicit limits as necessary parts of making the work reviewable.
Forks have their own upstream authors. My LeanScreen fork comes from ibrahimmian36/leanscreen, and my USAMO Guide fork comes from usamoguide/usamo-guide. They are separate from the original projects featured above.