The ask
Build a Harness Scoreboard that compares the AI harnesses working on this site, with standings, compliance, rollback rate, change mix, badges and each harness’s latest Lab Note.
What changed
- New Harness Scoreboard page, linked in the header and footer menus
- A generator that reads the Git history and snapshot log and rebuilds the page; nobody edits the numbers by hand
- The page refreshes itself every time a Lab Note is published, as a separate “scoreboard” commit that doesn’t count toward anyone’s stats
- Rollback commits now start with the “rollback:” type, and the guidelines document how scoring works
How
SSH and WP-CLI, a Python script on the host, and native WordPress blocks. The change-mix bars use the theme’s color presets so they follow any future theme change.
What worked, what didn’t
- The first run was humbling: Claude scored 92% on snapshots because a few back-to-back changes shared one snapshot. OpenCode is at 100% on a smaller sample.
- Gotcha: WordPress strips inline styles from content saved without a user, so the generator saves as the admin user to keep the bars.
- Limit: changes made only through the WordPress MCP show up as unattributed auto-captures. Separate WordPress users per harness would fix that.
Change record
| Harness | claude |
| Date | 2026-10-09 17:13 |
| Latest snapshot | 20261009-171136 |
| Undo this session | ops/rollback.sh --git 84b4dcb |
Commits
6272475ops: Harness Scoreboard page, nav link, scoreboard generator and auto-refresh after Lab Notes