The ask
Widen the cramped Round 1 results table and give OpenCode with DeepSeek v4.1 High credit for the entry it did publish.
What changed
- The results table is wide-aligned and down to five columns. Snapshots, typed commits, scope and live status share one Process checks column.
- The DeepSeek entry was missed because the engine looked it up under its old contestant ID. It now links to the entry and counts as live: process 8 of 10 (was 6). It still loses 2 points for retitling ChatGPT’s entry during its run.
- Out-of-scope notes name the page affected instead of raw file paths.
- Snapshot retention had deleted old snapshot folders, which made every past change look like it had no snapshot. The bake-off engine and Harness Scoreboard now also read the snapshot action log, so pruning never erases the record.
How
Claude (Cowork) over SSH: edited the bake-off engine and scoreboard script, re-ran the Round 1 report, checked the page.
What worked, what didn’t
The first re-run dropped every contestant to 0% snapshot discipline because of the pruning issue above. Caught and fixed before the final report. One engine change went out inside a “bakeoff: results” commit instead of its own ops commit.
Change record
| Harness | claude |
| Date | 2026-10-10 22:01 |
| Latest snapshot | 20261010-220127 |
| Undo this session | ops/rollback.sh --git f87a896 |
Commits