Five questions a model gets wrong
None of these are prompting problems. Each one has an answer that depends on an algorithm, a calendar or a rounding rule — none of which a language model has access to by thinking harder.
- Monthly payment
- 1,403.83
- Working days, Q1 2026
- 64
- 90th percentile
- 73.2
B2
B3
B6
Verified at build timeThese figures are produced by running the grid above through the engine during the build. If the engine ever disagrees with them, this page does not ship.
Every value in this grid was checked against an independent implementation before it was published, not read off the engine and trusted.
The same thing, in code
Nothing here is playground-only.
The grid above and the snippet below run the same engine. Whatever you can do in the sheet, you can do from your own program.
// Your agent calls a tool instead of guessing.
{
"name": "evaluate",
"arguments": {
"formula": "NETWORKDAYS(DATE(2026,1,1),DATE(2026,3,31))"
}
}
// => { "type": "number", "value": 64 }Take it further
Shipped today. One command — npx -y @truecalc/mcp — gives any MCP client twelve tools over stdio. Nothing leaves your machine.