The invoice that is off by a cent.
Same Harvard RCT: AI tutoring won, but only when it made you retrieve.
- The track
- Writes code
- The rep
- About thirty seconds
- Published
- September 2, 2026
The Exhibit
AI is not only a crutch in this newsletter's story. A Harvard RCT found an AI tutor matched or beat in-class active learning, conditional on the interaction being active: the tool prompted and questioned rather than just answering. The study.
It is a preprint, so hold it loosely. But the dividing line is the one you have read all series: retrieving versus receiving. The same model that will atrophy your skills if you let it autocomplete your thinking will sharpen them if you make it quiz you. The difference is entirely in how you drive it.
The study
Comprehension Debt
Comprehension debt: the gap between what gets shipped and what the team can still reason about.
So the move is not to close the model, it is to flip its role: instead of "write this for me," try "ask me three questions about this until I can defend it." You keep the speed and you keep the depth. The model will happily quiz you on exactly the kind of fact below, if you let it test you instead of tell you.
The idea
The Rep
Thirty seconds:
A billing feature stores money as a floating-point dollar amount. Tests pass, then finance finds invoices off by a cent and totals that will not reconcile. Why?
The answer
Binary floating point (IEEE 754) cannot represent most decimal fractions exactly. 0.1 + 0.2 does not equal 0.3; it comes out 0.30000000000000004, and that is not a glitch, it is the representation. Accumulate enough of those and rounding errors surface as off-by-a-cent invoices and totals that do not sum. The fix is to never store money as a float: use integer minor units (cents as an `int`/`bigint`), apply an explicit rounding rule on any multiply or divide (tax, splits, proration), and format to the major unit only at display. A fixed-precision decimal type (SQL `NUMERIC`/`DECIMAL`, or your language's decimal type) works too. Rounding patches treat the symptom; the type is the cause. This is one of those things AI will autocomplete wrong as cheerfully as right, which is exactly why it should be in your head and not the model's.
About thirty seconds