A few quizzes nobody studied for. 13 points.
The quizzed material hit 94. The same lessons, untested, stalled at 81.
- The track
- Writes code
- The rep
- About thirty seconds
- Published
- September 2, 2026
The Exhibit
Close the tab and try to rebuild the thing before you look. Recognizing it and reproducing it are not the same act, and the gap between them is what the study measured. A five-year classroom study (N=1,400+) changed one thing: brief, low-stakes quizzes inside normal lessons. Quizzed material scored 94% on the real exams; the same material, untested, scored 81%, and the gap survived into the finals. The study.
Students, not engineers, so treat the numbers as soft. The mechanism is the point: retrieval, pulling the answer out, is what builds durable memory. Recognition while re-reading feels like learning and mostly is not.
The study
Comprehension Debt
Comprehension debt: the gap between what gets shipped and what the team can still reason about.
Watching a walkthrough, skimming the AI's explanation, re-opening the file: all recognition. It feels like you know it. The proof you actually do is whether you can produce it cold, the day it breaks, with the docs closed. That is the only version of "I understand this" the on-call rotation cares about.
The idea
The Rep
Thirty seconds:
A nightly job calls a payment provider, sometimes times out, so you add a retry: on failure, call again. Customers start getting charged twice. What is wrong, and what makes the retry safe?
The answer
The first call may have succeeded server-side before the response timed out, so the retry charges a second time. A retry is only safe when the operation is idempotent, identical result whether it runs once or five times. The standard fix is an idempotency key: generate one per logical charge, send it on every attempt, and the provider collapses duplicates to a single charge. Retrying without that is not resilience, it is at-least-once delivery aimed at a side effect that should happen at-most-once.
About thirty seconds