How this was designed

Short version. Open a section for more.

What Anand asked for

How it works

Everything is done on the page. There is a small lab or map, then short questions in order. Nothing to buy or find.
Each browser gets its own set number from the checker. It picks the student’s own numbers. In the pendulum and salt labs it also picks hidden constants, so a friend’s readings do not fit.
Numbers are checked against the student’s own lab readings and against fresh cases made from their set.
Choices can block a top result but never grant one.
One sentence is checked by an AI for meaning. It can lower the result, never raise it.
The result is a signed code. Nothing is stored.

Why a student learns even with AI or a friend

The numbers come from your lab. The quickest way to get them is to run it and look.
A guess comes first, then a reading. The surprise is the learning.
Build tasks have many right answers. Friends compare different ones and talk about why both work.
Copying a friend’s numbers or an AI’s textbook answer gives wrong results for your set.

What was tested (15 Oct 2026)

Small test set written by us. No real students yet.

Test	Result
Answer rules: 5 tasks × all 90 set numbers	An honest sheet passes every rule for every set
Live checker: 30 labelled sheets, 5 tasks	All as expected: honest sheets Got it; shallow, wrong, keyword and injected sheets never Got it
Browser: all 5 tasks on a phone-width screen	Lab run, fill, check and result all work
Small model with no lab access	River, pendulum, salt and mixer: not Got it. Fence: Got it (see below)
Another student’s readings	Not accepted

Model: GPT-4.1-mini, for the one sentence only.

What it cannot do

It cannot stop a script or agent. One that drives the page can run the lab and pass. The lab still runs and every number is shown.
The fence task is pure maths. A model passes it with no lab. It teaches the idea only if the student does the building.
Re-checking many times can reveal which answers are right. A daily cap slows this.
A lightly reworded copy gets a different code.
A teacher who wants certainty can ask the 2-minute check on the teacher page.

Evidence behind the design

NCF-SE 2023 §3.3.7: homework should apply learning to different contexts. It warns about AI doing homework. PDF
Unrestricted GPT-4 raised practice scores 48% but lowered later exam scores 17% (Turkey, about 1,000 students). Bastani et al. 2025
Predicting before a demonstration improved understanding. Crouch et al. 2004
Homework links positively with achievement, more so in Grades 7–12 (US, flawed designs). Cooper et al. 2006
82% of rural 14–16-year-olds can use a smartphone; 57% of those used it for study last week. ASER 2024
Gap: CBSE’s own circulars on the new NCERT books could not be retrieved (404).

Checklist for a new task

Pilot plan

2–3 teachers, one class each, one task a week for 3 weeks.
Measure: the 2-minute check score, teacher minutes spent, and who could not finish and why.
Compare with a similar class given the usual worksheet.