Review complete. The core numbers were correct; I found and fixed three presentation/traceability issues, with no numeric changes.

## Review verdict

**Checks supporting the calculations** (all re-executed, all passed):
- Independent recomputation from the written cleaned CSVs matched hand-derived totals: Workshop 5 orders / 1,920.00 gross / 240.00 refunds / 1,680.00 net / 300.00 ads / 1,380.00 contribution / 336.00 per order; Course 4 / 900.00 / 60.00 / 840.00 / 150.00 / 690.00 / 210.00; combined 9 / 2,820.00 / 300.00 / 2,520.00 / 450.00 / 2,070.00 / 280.00.
- Row conservation re-asserted: sales 15 = 9 included + 1 duplicate + 1 superseded + 4 excluded; refunds 7 = 2+1+0+4; ads 7 = 5+1+0+1.
- Conflict handling verified against inputs/policy.md: S10 kept over S09 (latest updated_at, 180.00); S04/R03/A04 collapsed once each; refund ceilings hold (RF01 240 ≤ 480; RF02 60 ≤ 240).
- Requirement coverage vs PROMPT.md: all seven metrics × three bases, row-ID preservation, exceptions, reconciliation, contribution-vs-profit recommendation — nothing missing.
- Claims audit: every stated test maps to code actually run locally; no fabricated results.

## Corrections made

1. **Traceability**: kept rows now carry back-references to collapsed/superseded source rows (S03→S04, S10→S09, R02→R03, A03→A04), so every source row ID is traceable from the cleaned files themselves, not only exceptions.csv.
2. **Scoped provisional claim**: summary now states the upper-bound caveat applies to Course and combined only (Workshop unaffected by RF05), and adds combined bounds under the paid-gross ceiling: net revenue 2,160.00–2,520.00, contribution 1,710.00–2,070.00.
3. **Wording**: removed "contributes 0" for RF05 (risked implying a missing amount was zeroed, which policy forbids); now "excluded from the calculable totals; unknown amount not treated as zero".

Post-correction, the extended verifier confirmed all non-note fields, inclusion flags, totals and exceptions.csv are unchanged versus the preserved original ("REVIEW CHECKS PASSED", "ALL CHECKS PASSED").

## Revised deliverables

- `output/cleaned-sales.csv`, `output/cleaned-refunds.csv`, `output/cleaned-ads.csv` (notes enriched; data unchanged)
- `output/exceptions.csv` (unchanged, 13 rows)
- `output/summary.md` (revised wording/bounds)
- `output/reconcile.py` (corrected, reproducible generator)
- `output/CHANGELOG.md` (review checks + change log)
- `output/original_submission/` — all six original files preserved unchanged; `inputs/` untouched

**Remaining limitations** (unchanged): results provisional pending RF05's missing amount (R06); RF04's order N9999 doesn't exist in inputs; RF05's ceiling uncheckable; no cost data, so contribution ≠ profit.

**Estimated human finish time**: ~20–30 minutes (obtain RF05 amount, investigate N9999, re-run) — an estimate, not a measured result.