Manu Kapur’s 2008 experiments with middle schoolers in Singapore compared two approaches. Group A received instruction first, then solved problems. Group B attempted the problems before any instruction, inevitably failing — then received the same instruction Group A had. On delayed transfer tasks, Group B outperformed substantially.
The explanation: failed attempts activate the learner’s existing mental models, expose their edges, and create specific questions that instruction can then answer. Instruction delivered onto unprepared ground is absorbed passively; delivered onto a mind that has just struggled, it lands as revelation.
Fluera’s Step 2 — reconstruct the canvas from memory before consulting anything — is a productive failure machine. The red nodes where you fail are exactly where the Socratic AI directs its attention in Step 3.