The reforecasts got the full year right on costs but kept forecasting revenue too high. On paper the best version is the 9+3, with a full-year revenue variance of $3K, or 0.02%, and a MAPE of 0.63%. It also had only three months left to forecast, so it's the shortest-horizon version rather than proof the forecasting got better.

**Revenue: every version except 9+3 forecast too high**

| Version | Full-year variance | MAPE | Bias |
| --- | --- | --- | --- |
| Budget | $362K, or 3.2% | 3.7% | 3.2% |
| 3+9 | $177K, or 1.6% | 3.3% | 2.1% |
| 6+6 | $158K, or 1.4% | 2.9% | 2.8% |
| 9+3 | $3K, or 0.02% | 0.63% | 0.10% |

- **The upward revenue bias held for three versions.** The 6+6 still had a 2.8% bias after six closed months, so reforecasting didn't fix the optimism.
- **Revenue had a few big single-month misses:**
  - The 6+6 missed 2025-09 by $151K, or 15.0%, against a MAPE of 2.9%. That's one large miss; it's worth asking what that forecast assumed for September.
  - The Budget missed 2025-07 by $132K, or 15.0%, and the 3+9 missed the same month by $80K, or 9.1%. Both carried a July assumption that didn't happen.
- **Operating expenses got more accurate with each version.** MAPE fell from 4.1% in the Budget to 0.30% in the 9+3.
- **The Budget's full-year opex number is misleading.** Its bias was only (0.03%), but its MAPE was 4.1%, so large monthly misses cancelled out. The worst was 2025-05, which was over-forecast by $32K, or 15.7%.

All figures come from the analysis script. It confirmed that every closed month matches the P&L actual and that the Budget version matches the P&L budget.

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