File 05 · AI hackathon · Casablanca · Sep 2026
Wolf Station
I didn’t use AI to read the spreadsheet. I used it to write the program that reads the spreadsheet. Then that program kept working for free.
- 4/4late files exact to the cent against the answer keyMeasured
- 13 ms192/192 lines, 8× larger file, zero tokensMeasured
- €75.5Merror a naive import makes, reading forints as eurosMeasured
A supplier file arrives late, in a new layout and a new language, and has to change a buyer’s decision safely. The fictional company is a carmaker and the data is synthetic.
Letting the AI write the rows was usually right, but one run invented a row and another silently returned 24 of 192 lines. So the AI writes a small JSON recipe instead, plain code runs it, and an auditor checks it against the sheet’s own arithmetic, never the answer key.
Six agents with their own identity and memory can explain anything and approve nothing. Every file, recipe, decision and correction becomes a note in a graph, written from an append-only log.
What it refuses to doStaying confident when the ground moves. New evidence turns an approved decision stale, unclear rows stop and ask a person, and a hidden instruction in an invoice is ignored.
How it runs
- RecipeThe AI writes an importer: which column means what, which totals must balance.
- RunPlain code executes the recipe. No AI.
- AuditChecked against the sheet’s own arithmetic.
- ApproveA person approves the recipe once.
- ReuseEvery later file in that layout imports at zero AI cost.
Honest caveatFictional company and synthetic data supplied by the challenge. All numbers re-run without AI.
Next fileVox →