# Commentary review batch — 2026-06-27 (for Martin)

Ten files to review, in English and Swedish, for five races: the four you've already commented on
(the May-24 GSYS replays) plus one clean real-course example.

| File | Race | Course |
|------|------|--------|
| `gsys01_en.csv` / `gsys01_sv.csv` | 2026-05-22 15:09 | real V2 course (definitive winner) |
| `gsys06_en.csv` / `gsys06_sv.csv` | 2026-05-24 10:46 | reconstructed (preliminary) |
| `gsys07_en.csv` / `gsys07_sv.csv` | 2026-05-24 11:34 | reconstructed (preliminary) |
| `gsys08_en.csv` / `gsys08_sv.csv` | 2026-05-24 13:11 | reconstructed (preliminary) |
| `gsys09_en.csv` / `gsys09_sv.csv` | 2026-05-24 14:47 | reconstructed (preliminary) |

## What's in each row

One row per spoken line, with the context to judge it **without watching the render**:

- **clock** — race time vs the gun (`-1:30` = 90 s before the start)
- **role** — Caller (facts) or Analyst (colour/tactics)
- **type** — the kind of call (phase, play, row, margin, winner, boat_intro…)
- **leader / leg / about** — who leads, the leg's point of sail, and the boats the line is about
- **text** — the line as it is spoken today (the rule-based engine)
- **ai_suggestion** — *new:* for the colour/tactics lines, what the **Level-4 AI author** would say
  instead (already fact-checked by the safety validator — it can't invent a sail number, a margin, the
  wrong leader, or a winner before the finish). Blank for factual lines (the AI never touches facts).

## What we'd love you to fill in (three columns)

- **verdict** — `ok` · `wrong` (factually wrong) · `dull` (filler/irrelevant) · `irrelevant`
- **better** — what a good caller would have said here (your gold line). If the **ai_suggestion** is
  already good, just write "ai" — that's the strongest signal of all.
- **tag** — optional reason: `fact` · `repeat` · `timing` · `tone` · `names` · `other`

Open in Excel (they're UTF-8 with a BOM, so Swedish/Chinese render correctly). When you send them back,
we run `review_commentary.py --ingest <file>` and it prints your % ok/wrong/dull/irrelevant per race and
builds the dataset that trains and grades the AI author. Every line you mark is wasted on nothing — it's
exactly the gold the AI has to learn and be measured against.

(Tool: `demo/export/review_commentary.py`; workflow: `demo/data/reviews/README.md`.)