I've written the review to `output/register-note.md`, and the script's figures are in `output/register.json`. The script's counts reconcile.

**Result:** 4 of the 20 assumptions are flagged.
- **Stale review:** A-09 Inventory turns (Tomas Herrera) and A-13 Line of credit interest rate (Dana Whitfield).
- **No owner:** A-16 Merit increase.
- **Changed with no source:** A-04 Average headcount went from 23.00 to 23.75 in the Current version with no source recorded. This one matters most because it feeds the Headcount driver, so it moves opex and EBITDA. Elena Brandt owns it.

5 values changed in total. The other 4 changes (A-01, A-03, A-07, A-08) each have a source recorded.

**Before you rely on this, check the review date.** The script measures staleness against a default date of 2026-01-01, not today (2026-10-04). The skill says to change that date only if you give a different one, so I left it. The cutoff for "stale" is 180 days, so against today, every assumption last reviewed before about early April 2026 would count as stale. That probably includes every row in the register, so the true stale count is likely well above 2. Say the word and I'll rerun with `--as-of 2026-10-04`.

**The script rounds two percentages in the review.** The review has to copy the script's text exactly, so it shows the rounded values:
- A-13's rate is stored as 0.0725 but shows as "7.3%".
- A-03's new margin is stored as 0.2655 but shows as "26.6%".

If the review needs the exact rates, it's worth fixing the script's rounding.