# 2.4 facts the English chapter may use

**Setup.**
- 28 W+L memos, raw Grok transcripts, temperature 0. `names.md` is attached in both conditions.
- Ask: "Clean up this voice memo transcript. Output only the cleaned-up full text."
- The skill (`cleanup_skill.md`) has 5 rules:
  - keep the speaker's "I mean";
  - fix names per `names.md`;
  - write numbers and times as digits with a.m./p.m.;
  - don't merge or summarize;
  - keep the length within 85–115% in words.
- Pass = 3 checks:
  - every "I mean" in the transcript is kept;
  - the numbers rule holds (5 targets: W02 "ten in the morning", W08 "eight at night", W12 "ten minute", L02 "around nine", L05 "forty cents");
  - the length is in band.
- 15 of 28 memos have something to check. A model that changed nothing would pass 23/28.

**Results.**

| | pass | "I mean" kept | numbers | length in band |
|---|---|---|---|---|
| deepseek-v4.1-flash, no skill | 22/28 | 10/11 | 0/5 | 28/28 |
| deepseek-v4.1-flash, with skill | 28/28 | 11/11 | 5/5 | 28/28 |
| gemini-3.8-flash, no skill | 18/28 | 5/11 | 0/5 | 28/28 |
| gemini-3.8-flash, with skill | 28/28 | 11/11 | 5/5 | 28/28 |

- Gemini deleted "I mean" in W03, W15, W20, L03, L05 and L07.
- Default number styles were gemini "10:00 AM" / "8:00 PM" and deepseek "8 PM".
- Mean word similarity to the transcript was 0.95 (gemini) and 0.97 (deepseek). No preambles, no summaries.

**Quotes (verbatim).**
- L03 Grok: "Hard-boiled eggs are fine too. I mean, just peel them first so there's no shell left on."
- L03 deepseek, no skill: "Hard-boiled eggs are fine too — just peel them first so there's no shell left on."
- W02 gemini, no skill: "Tomorrow at 10:00 AM". With the skill: "tomorrow at 10 a.m."
- W01: only names and one sentence break changed (see README).

**Do not write** that English models rewrite into memo-speak or summarize by default.