The review found no calculation errors; all the figures stand unchanged. I made five fixes, mainly to file usability and two overstated claims in my first handoff. The original submission is saved unchanged in `output/original-submission-v1/`.

**Results (still provisional because refund RF05 has no amount):**

| | Workshop | Course | Combined |
|---|---:|---:|---:|
| Net revenue (CAD) | 1,680.00 | 840.00 | 2,520.00 |
| Contribution after advertising | 1,380.00 | 690.00 | 2,070.00 |
| Net revenue per paid order | 336.00 | 210.00 | 280.00 |

**Changes made:**
1. **`exceptions.csv`:** added `category` and `needs_follow_up` columns. The two items that need action now stand out: RF04 (order N9999 isn't in the sales file) and RF05 (no amount). The other 11 rows are routine exclusions under the policy.
2. **`summary.md`, wrong label:** the RF04 assumption was headed "out-of-period order"; it now reads "unknown order", which matches the data.
3. **`summary.md`, additions:**
   - a short guide to reading the output files;
   - a note that Workshop's higher per-order figure comes largely from one 720.00 order (N1010);
   - a list of the checks that were actually run.
4. **Unrecorded hand check:** my first handoff said I'd worked out the totals by hand separately. I did that only in my head, so there's no record of it. I replaced it with `verify.py`, which recomputes every total from the cleaned CSVs and can be rerun.
5. **Unchecked claim about inputs:** I'd also said the raw input files were unchanged without checking. The scripts only read them, and the write tool can't edit inputs. I've now recorded a fingerprint of each file, but there's no earlier copy to compare against.

**Checks actually run:**
- **`reconcile.py`:** included + excluded rows equal the raw rows in each file, and contribution = gross − refunds − ads. Both passed.
- **`verify.py`:**
  - every figure in the results table matched;
  - included order, refund and ad IDs are each unique;
  - all 13 excluded rows appear in `exceptions.csv`;
  - no applied refund exceeds its order's gross.
- **Manual review:** I read the regenerated `exceptions.csv` and confirmed the new columns. I also confirmed that the date, latest-record, duplicate, pending and period rules were applied as the policy requires.

**Remaining limitations:** the RF05 amount and order N9999 can't be resolved without payment-processor records. I can't directly prove the inputs are unchanged since the first run, because there's no earlier copy to compare.

**Estimated human time to finish:** about 15–25 minutes to get RF05 and RF04 resolved, rerun `reconcile.py` and `verify.py`, and update the summary. This is a guess, not a measurement.

Files are in `output/`:
- `summary.md`
- `cleaned-sales.csv`
- `cleaned-refunds.csv`
- `cleaned-ads.csv`
- `exceptions.csv`
- `reconcile.py`
- `verify.py`
- `CHANGELOG.md`
- `original-submission-v1/` (original submission)