# Change log — review of Task 2 submission

The recommendation (O2) and all six landed totals are unchanged. The original files are preserved unmodified in `output/original_submission/`.

## Corrections

1. **Discrete-GPU assumption recorded.** The packet never states that any GPU is "discrete". The original marked this requirement as met without saying so. `recommendation.md` now states the assumption, and `comparison.csv` has a `discrete_gpu_ok` column set to `ASSUMED` for every row.
2. **Citations tightened.** The original cited filenames without the `inputs/` location and did not quote section headings. The revision adds a source key with exact filenames and the exact heading of each markdown source; `comparison.csv` sources now carry the `inputs/` prefix.
3. **Unknowns made consistent.** The original listed "stock not stated" only against O2 and "None" for the others. Stock and order cut-off time are unstated for every offer; all rows now say so, and the limitation is in `recommendation.md`.
4. **O6 traceability.** `comparison.csv` now has `delivery_date_S2` beside `delivery_date_effective`, so the superseded October 23 date and the S3 date of October 27 are both visible.
5. **S3 timestamp remark removed.** The original implied a tension between S3's 08:45 time and the register's 09:00 capture time. The 09:00 time is when the packet was captured, and S2 carries no timestamp of its own, so there is no conflict to resolve. The precedence of S3 rests on the register and S3's own wording.
6. **Word-count margin.** The original was 649 by whitespace count against a 650 limit. The revision is 645 by the same count (564 tokens that contain a letter or digit); the O2 requirement table became a bullet list to make room for items 1 and 2.

## Checked and left unchanged

- Landed totals, tax, budget margins and pass/fail flags (recomputed from the input CSVs).
- O2 as the only qualifying offer.
- Treatment of the O3 and O4 listing titles, the O5 estimate and the O6 correction.
- Fallback approach.
