# Deviations from RULE.txt

1. **Aborted first launch (2026-10-20 ~09:56Z).** The six lanes were started from zsh, which passed
   `"computer-mcp-0112 t1"` as one arm name. Every run failed at arm setup with `FileNotFoundError:
   …/arms/computer-mcp-0112 t1.txt`, before Claude Code started, so no agent saw any task. Those
   rows are kept in `results-aborted-zsh.txt` and are not scored. The lanes were relaunched from bash.
2. **Two debug runs outside the protocol.** While finding (1), task 1 (`libreoffice_calc d681960f`)
   was run twice with our arm (`trial dbg` and `dbg2`), both reward 1. They are not part of the
   three scored trials and are not counted.
3. **Turn cap.** RULE.txt names OSWorld's evaluator as the score and does not mention the 100-turn cap; a
   capped run is never evaluated (`reward=ERR` in results.txt, an empty reward and `subtype error_max_turns` in runs.csv). The headline follows the rule as
   written and leaves those runs out (`score.txt`: 51/58 vs 49/58). At about 09:59 UTC, after runs began, I also
   wrote down scoring them 0; that is reported only as a secondary view (`score-capped-as-zero.txt`: 51/60 vs
   49/60). The cap hit 2 runs per tool, all on task 5 (`multi_apps 69acbb55`).