Receipts
Pull an AI answer down to see the rows behind each number, and catch the one that doesn't add up.
Pull the answer down or push it up. Point at a number to see it add up.
How much overtime did Support log last month, and who drove it?
Support logged of overtime in October across , costing . drove the most, with .
| Employee | Team | Hours | Pay |
|---|---|---|---|
| Ana Ruiz | Tier 1 | 4 | $200 |
| Dana Okafor | Tier 2 | 9 | $450 |
| Leo Brandt | Tier 1 | 6 | $300 |
| Mia Castell | Tier 2 | 5 | $250 |
| Priya Natarajan | Tier 3 | 22 | $1,100 |
| Sam Whitlock | Tier 1 | 5 | $250 |
| Sam Whitlock | Tier 2 | 3 | $150 |
| Tomás Ferreira | Tier 3 | 10 | $500 |
| Total | 64 | $3,200 |
One number in this answer is wrong. Try to find it before you read on.
Pull the answer down to bring up the overtime log it came from. Point at any figure and the cells behind it light up and add up. 64 hours checks out. So does $3,200.
The bad one is “8 people.” Sam Whitlock is in the log twice, once under Tier 1 and once under Tier 2, and the answer counted him twice. That’s the mistake I worry about most from AI summaries: plausible, off by one, and invisible unless the arithmetic is on screen.
I’ve argued for this in AI needs receipts: the evidence should sit at the level of each claim, one figure and its cells. Here it’s one pull away. The part I’d still have to solve for a real product is scale. Eight rows fit under a card. Eight thousand don’t, and I don’t have a good answer for that yet.