The 9+3 reforecast landed closest. It missed full-year revenue by $3K, or 0.02%, with a MAPE of 0.63%. It was also the closest version on operating expenses. The script's checks passed: every version's closed months matched the P&L actuals.

**Revenue: we ran high in every version.** All four versions forecast more revenue than came in.

| Version | Variance | Bias |
| --- | --- | --- |
| Budget | $362K, or 3.2% | 3.2% |
| 3+9 | $177K, or 1.6% | 2.1% |
| 6+6 | $158K, or 1.4% | 2.8% |
| 9+3 | $3K, or 0.02% | 0.10% |

- **The mid-year reforecast stayed optimistic.** 6+6 still had a 2.8% bias with six months already closed.
- **Two months account for the largest misses:**
  - **July 2025:** the Budget over-forecast it by $132K, or 15.0%, and 3+9 by $80K, or 9.1%.
  - **September 2025:** 6+6 forecast $1,156K against an actual $1,005K, a miss of $151K, or 15.0%.

  It's worth checking what those forecasts assumed for July and September.
- **9+3's accuracy partly reflects timing.** It had only three months left to forecast, so its low error doesn't by itself show that our forecasting got better.

**Operating expenses: no consistent direction.** Bias was slightly negative in Budget, 3+9 and 6+6, and slightly positive in 9+3. The largest was (0.74%) in 3+9. Every version landed within $14K of the $2,527K actual.

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