Each reforecast got closer to the actual year than the last, but most of that came from having fewer months left to forecast. Revenue was forecast too high all year. The script's checks passed: every closed month matches the P&L actual, and the Budget version matches the P&L budget.

**Revenue**
- **The 9+3 reforecast was the most accurate.** It missed the full year by $3K, or 0.02%, and its average monthly miss was 0.63%. But it had only 3 months left to forecast, so it shows little about forecasting skill.
- **Revenue was too high in every earlier version.** The budget forecast too high by 3.2% on balance, the 3+9 version by 2.1% and the 6+6 version by 2.8%. Two reforecasts didn't fix it. The data doesn't show which assumption caused it; a customer-level or driver-level forecast would answer that.
- **6+6 had one large miss.** It overshot September by $151K, or 15.0%, against an average monthly miss of 2.9%. The budget had a similar single miss in July: $132K, or 15.0%. Ask the planners what each version assumed for those months.

**Operating expenses**
- **The budget's full-year total was close, but its months were not.** It missed the year by just ($649), or (0.03%), yet its average monthly miss was 4.1%, because over- and under-forecasts cancelled out. The worst month was May, at $32K, or 15.7%. The weakness was spreading costs across the months, not the annual total.
- **Reforecasts improved steadily.** The average monthly miss was 1.8% for 3+9, 0.77% for 6+6 and 0.30% for 9+3.

**Which version is the fair test:** 6+6, because it still had six open months. It forecast operating expenses well, but it still overstated revenue. Next year, I'd treat early revenue reforecasts as likely too high until that pattern is explained.

Files are in `output/`:
- `accuracy-note.md`
- `accuracy.json`