Excel — the brief, the fixture and the scoring

Sent as one message with orders-q2.xlsx attached through each tool’s own file upload, on the free tier, in a fresh chat (Claude: incognito). Nine tools, same file, same day. If a tool cannot take an .xlsx, the sheet is pasted as CSV text and that is noted.

I run a small online shop in Portugal. Attached is my order export for April to June 2026 - it’s messy. Please:

1. Tell me what’s wrong with the data and fix it. Give me the cleaned sheet back as an .xlsx.
2. Revenue per month and per category, and the product with the most units sold.
3. List the duplicate orders and any order where Total doesn’t match Qty × Unit Price.
4. One Excel formula I can paste to get revenue per month from the cleaned sheet, which keeps working when I add rows.

Keep it short.

The fixture

orders-q2.xlsx, one sheet “Orders”, 124 rows: 120 orders plus 4 exact duplicates. Built by make_fixture.py (seeded, reproducible); the exact answers are in ground-truth.txt. Eight faults are planted:

Fault	Rows	What a careful reader does
Exact duplicate orders	4	drops them
Total ≠ Qty × Unit Price	5	flags them; revenue uses Qty × Unit Price
Qty blank	3	derives Qty from Total ÷ Unit Price
Region blank	2	flags, cannot be derived
Category blank	8	fills from the product name
Unit Price as text, “€12,50”	18	converts to a number
Product name case and spacing variants	18	normalises to one spelling per product
Dates in three formats: real dates, dd/mm/yyyy text, Apr 5, 2026 text	30 + 12	parses all as dates; the shop is in Portugal so 03/04/2026 is 3 April

The date trap has a clear tell: reading dd/mm as mm/dd produces orders in July to December and 20 invalid dates. ground-truth.txt also lists the totals a tool gets if it sums the Total column as-is, or deduplicates but keeps the wrong totals.

Scoring

• Faults found - how many of the eight fault classes the tool names, of 8.
• File - an .xlsx comes back, opens, and is actually cleaned: 120 rows, real dates, numeric prices, one spelling per product, categories filled. Checked with openpyxl.
• Months - revenue for April, May, June within €0.01 of the truth, of 3. A stated different basis (for example “sum of the Total column as given”) is recorded but not counted as right, because the brief asks for the cleaned data.
• Categories - of 3. Top product - Vinho Verde 750ml, 42 units, right or wrong.
• Duplicates - of 4, with any false positives noted. Mismatches - of 5.
• Formula - works on the cleaned sheet, and keeps working when rows are appended (whole-column or Table references, not a fixed range). Judged by reading it and, where the logic is not obvious, re-implementing it.
• Invented facts - anything stated about the shop, the customers or the numbers that the file does not support.
• Time to answer, whether the free tier refused or truncated, and whether the answer is in English.

Fixtures for the Excel test

• orders-q2.xlsx - the messy order export sent to every tool. Synthetic: a fictional Portuguese food-and-homeware shop, invented customer names, no real people or businesses. Built by make_fixture.py.
• make_fixture.py - generates the workbook and prints the ground truth; seeded so it reproduces exactly.
• ground-truth.txt - the right answers, plus the alternates a tool produces if it skips a cleaning step.
• brief.md - the message sent, the fault list and the scoring rules.
• check_returned.py - opens a returned workbook and compares it with the ground truth (row count, duplicate IDs, date and price types, product spellings, blanks, monthly totals).
• returned/ - the cleaned workbooks the tools sent back, named by tool. Claude, Grok, Copilot and Perplexity returned one; the others did not.

Cleanup afterwards: Copilot’s composer said a copy of the upload goes to the personal OneDrive (“При отправке с устройства копия будет отправлена в OneDrive (личный)”) - delete orders-q2.xlsx there. Nothing else persists: one upload per tool, no memory instructions, no automations.