The reforecasts got more accurate as the year went on, but mostly because each later version had fewer months left to forecast. Revenue forecasts also came in too high in every version before 9+3.

The figures below come from the review script, and its checks passed. Closed months in every version match the P&L actuals, and the Budget version matches the P&L budget.

**Revenue**
- **Best version:** 9+3 was the most accurate. It missed the full year by $3K, or 0.02%, and its typical monthly miss (MAPE) was 0.63%. But nine of its twelve months were already actuals, so that mostly reflects closed months, not better forecasting.
- **Budget, 3+9 and 6+6** all forecast too high on balance, by 3.2%, 2.1% and 2.8%.
- **Two large single-month misses:**
  - **Budget, July 2025:** over by $132K, or 15.0%. The 3+9 version was also over that month, by $80K, or 9.1%.
  - **6+6, September 2025:** over by $151K, or 15.0%, against a typical monthly miss of 2.9%.

  The data doesn't show what the forecasts assumed for those months. The pipeline or phasing assumptions behind them would answer that.

**Operating expenses**
- **Budget looks better than it was:** it landed within ($649), or (0.03%), for the full year. But its typical monthly miss was 4.1%, the highest of any version, because monthly misses in opposite directions cancelled out. Its worst month was May 2025, over by $32K, or 15.7%.
- **6+6 is the best sign of a sound process:** its typical monthly miss was 0.77% with six months still to forecast.

**What to rely on:** for opex, 6+6. For revenue, no version shows a reliable monthly process yet. Check the phasing of big months like July and September before anchoring on any version.

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