[ research & tools ]
what does a score actually measure? what evidence makes a result worth trusting? a few projects exploring those questions.
math reward hacking
saved-output case study
when a reward improves, did the answer improve? follow one saved response through different scoring rules to see how completion limits and answer extraction change the result.
scope & limitations
measurement choices materially affect the comparison. the study does not establish consistent reward hacking. three training-seed pairs limit generalization; model weights are not distributed. the repository includes saved completions, analyses, and offline verification.
reviewops
local prototype
what was reviewed, and which evidence did it use? a local workspace for experiment handoffs: compare submissions, keep review notes, and preserve earlier evidence and decisions.
scope & limitations
startup cases are synthetic. file hashes check content identity, not scientific correctness. the app runs locally with python and includes offline backup and restore. researcher adoption and time savings have not been measured.
research design · documentation only
does an evaluator's verdict change when a statement's presentation changes but its meaning does not? a proposed study with unchanged-input controls and a plan for inspectable evidence.
scope & limitations
the public repository is a documentation scaffold. no runnable evaluator, proof certificates, dataset, tests, or experimental findings are released there.