A small internal tool for logging and comparing test runs across different Claude models (Sonnet 5, Fable 5, Mythos 5, Opus 4.8, Haiku 4.5, etc.) on tax document PDF and Excel input tasks — timing, output quality, and pass/fail, all in one place.
- Log a run — task/document tested, model, reasoning effort, a built-in stopwatch (or manual override) for timing, a 1–5 quality score, and a Pass/Partial/Fail result.
- All Runs — every entry you've logged, sortable by any column, deletable.
- Report — auto-generated comparison by model: run count, average/fastest/ slowest duration, average quality, pass rate, plus a one-line callout on what's fastest and what scores best. One click copies a plain-text version of the report for pasting into an email or notes.
- Make sure the tracker file is named exactly
index.htmland sits at the root of the branch/folder you're serving from. - In Settings → Pages, set the source to the branch and folder (root,
unless you're using
/docs). - Add an empty file named
.nojekyllat the repo root — this skips GitHub's default Jekyll build, which otherwise can interfere with a plain HTML file. - Give it a minute to rebuild, then hard-refresh the page.
This version stores logged runs in the browser's local storage — not in the repo and not synced anywhere. That means:
- Data lives in whichever browser/device you use to log it. Same browser next time to see your history.
- Clearing that browser's site data will clear the log.
- Nothing you log here gets committed to GitHub — the repo only holds the blank tool itself.
Internal tool. Not intended for public/client use.