AtomicRepsAfter their coding agent finishes, the student answers a question about that code from memory. What they retrieve, they strengthen.
the reminders list shows other users' reminders after the first request. fix
⏺ Read(app/reminders.py)
⎿ Read 61 lines
⏺ Grep(pattern: "reminders=", path: app)
⎿ 2 matches
⏺ Update(app/reminders.py)
⎿ Updated 1 file, +3 -1
⏺ Fixed it in app/reminders.py. Want me to check the other handlers for the same thing?
Atomic Reps · PythonREP 01
reminders=None. What did the old signature do the second time it was called without a list?def add_reminder(text, reminders=[]):
reminders.append(text)
return remindersYour student's terminal, week six. Pick a letter.
Working code used to be evidence that a student had worked through the problem. With coding agents, those two things have come apart. What happens between the lab and the hall exam has been measured. Atomic Reps watches no screen and runs no detector. It asks the student a question from memory, and the answer is the evidence.
After the task, roughly 50 against 67. The AI group was slightly faster, but not significantly so. Fifty-two developers whose coding AI assisted, pre-registered trial.
Reaching a working solution with AI and, in the authors' words, “an illusion of competence”. In the same lab's pre-AI study, 11 of 31 could not finish, and every one of them knew it.
In 2025. One response has been more invigilated hall exams: an unassisted signal, but one that arrives at the end of the term.
“GenAI is not cognitive offloading. It’s outsourcing.”
Which concepts the class missed this week, with the question beside each one, so the next lecture starts from what the class got wrong rather than what the slides assumed.
38 of 41 answered. 3 students paused themselves.
def add_reminder(text, reminders=[]):
reminders.append(text)
return remindersHalf the class missed it. Start with A and D.
Ranked by how many of the 38 answers got it wrong, against the same rep everywhere else it has been served.
What did reminders=[] do the second time the function ran?
A default is built once, when def runs, and it belongs to the function. Every later call without a list appends to that same one.
After b = a and b.append(4), what does a hold?
b = a binds a second name to the same list. a[:] builds a new one.
You wrote names = names.sort(). What is in names now?
list.sort() orders the list in place and hands back nothing. sorted() leaves the original alone and returns the new list.
Two lists hold the same numbers. What does is say about them?
is asks whether the two names point at one object. == asks whether the two objects hold the same thing.
You have t = (1, [2, 3]). Does t[1].append(4) raise?
The tuple fixes which objects it holds, not what those objects contain. The list inside it can still grow.
Each one prints with its reason and its source.
Sixteen of the nineteen wrong answers say the default is rebuilt on each call.
The room answers, then the split lands on screen with no names, then the reason.
To the fourteen who missed it, nine days after they first met it.
Thin against the rest of your term. We write the missing ones from your own syllabus, in your licence year.
Every question is reviewed by a person and fixed weeks before your term starts, so nothing is invented while a student is answering. A class of forty meets the same question the same way, and you veto anything you have not taught.
The questions are about the code the student just shipped: the default argument they left mutable, the list they aliased, the exception they swallowed. A few seconds with their own work, each time.
Fixed rules choose each next question: the week's topics, the level band, nothing seen recently. Every question is fixed before it is served, and no model grades a student. What you read on the class screen is arithmetic over answers.
Questions written against your course plan, in your words, reviewed by people and folded into the corpus. Included in your licence year. Write to us.
Built from official product documentation and from openly licensed university and interview preparation material, adapted for retrieval practice.
Being asked to recall beats reading it again, and the lead widens with every day between the lesson and the exam. So the student is asked, not shown. One question about the code they just shipped, answered from memory.
Recalled one week later · equal time on task
Five minutes after studying, the rereaders were ahead. A week later the recallers were, and an exam is weeks away. Lab study, prose passages.
For the material that brief low-stakes questions had covered, 94% against 81%.
48,478 students, pooled effect g = 0.499, and five publication-bias checks left it near-clean.
“I’m all about adding friction to force reflection.”
Every number links to the page that grades it. The full reading is on the retrieval practice page and the cognitive debt page.
Atomic Reps follows the course. The course never has to follow Atomic Reps. It is built to sit inside a department's year, its exam calendar and its procurement rules, and nothing is asked before it has been taught.
A twelve-week term as the console plans it. Example plan; the marked weeks are the product's.
One invoice, nothing above it to buy. The department knows the figure before it opens a procurement file.
Seats belong to the school rather than to a student or a course, so a department that finishes one class moves the same seats into the next one.
Excluding VAT · invoiced by Atomic Reps AB · SEK with 25 percent moms in Sweden · reverse charge for other EU institutions · no auto-renew · price rises capped at 3 percent a year.
Start in one computer science course. One licence year, invoiced once, and you have the class screen from the first rep.
Your students keep the agent. You get the week the class actually had.
AtomicRepsYour students keep the agent and the tools they will use outside school. You get the one thing the exam cannot give you until it is too late: what the class can still explain without them.
One school, one invoice.