← all simulations
Cookbook 10 The Python 10_bookkeeping.py

A month of transactions

This is the one where calibration stops being a talking point. What matters is not how often the model is right — it is whether it knows which rows it is unsure about. If it does, the bookkeeper reads sixty rows instead of eight hundred, and the eight hundred are not a leap of faith. Drag the floor and watch both numbers move.

DateWhat the bank saysAmount Filed asSure
Press Play.

The floor — set it from the table on the right, not from taste

Rows
0
on the statement
Filed
0
nobody read them
Review queue
0
the bookkeeper’s morning
Filed correctly
vs. the human ledger

Is the confidence number load-bearing? — last year’s ledger, re-run

Take a few hundred rows you already categorised by hand, run them through, and bucket the results by the confidence that came back. Accuracy climbing with confidence is what makes a floor meaningful. A flat table means the number is decoration and nothing should be filed automatically — which is a thing you can find out in an afternoon, before you trust it with anything.