90 second guided case
A wrong answer has a rule inside it
52 − 28 = 34
Play one learner’s reasoning. GLITCH finds the rule, turns it into a boss, and the learner breaks it with their own counterexample.
Same engine, three learners
Three learners answer the same problems, each following a different rule. Nothing here is scripted: the page runs the real diagnosis on their steps as it loads.
Learner A
52−28→34 31−15→26
Free Ten wakes up after 2 problems.
Learner B
52−28→36 31−15→24
Flip Flop wakes up after 2 problems.
Learner C
52−28→24 31−15→16 64−37→27 83−46→37 75−18→57 42−19→23
No boss after 6 problems. No rule to fight.
Play it yourself: regroup correctly and no boss appears; take the smaller digit from the bigger and Flip Flop wakes instead.
AI where language is fuzzy. Code where math must be true.
The model proposes rules as small programs and reads the learner’s explanation. Deterministic code runs every rule, checks every counterexample and decides every win.