Attention within a lesson. A vigilance-decrement toy model: a(t) = a∞ + (1 − a∞) e−t/τ, restarting at 1 after every active break. The mean over a lesson of length L has the closed form 〈a〉 = a∞ + (1 − a∞)(τ/L)(1 − e−L/τ) (periodic version with breaks). Classroom studies disagree about how fast attention really fades (often quoted as “10 minutes” but poorly supported), so τ and a∞ are sliders, not facts.
Forgetting. Retention after a delay is R = e−t/S where S is the memory stability in days (Ebbinghaus-style exponential). The stability a lesson creates is taken as S0 = Sbase · 〈a〉: better attended teaching gives a stronger trace. Sbase is an illustrative number, not a measured constant.
Retrieval practice and spacing. At each review, success resets R to 1 and strengthens the trace: S ← S (1 + h + g (1 − R)) with h = 0.3 (mere re-exposure) and a gain g that grows when the answer was harder to recall (the “desirable difficulty” idea). Cramming right after class has R ≈ 1, so it earns only h; spaced reviews happen at lower R and earn more. The spacing effect and the testing effect are well replicated; the exact functional form here is a simplification.
Virtual cohort. The 20 visible students have individual stabilities S·exp(0.35 z) (log-normal spread). The cohort score at the test is 20 000 hidden Bernoulli trials with success probability R of the typical learner.
Checks. (1) The mean attention closed form against a midpoint integration of the raw definition. (2) The event-recursion value of R at the test against a separate RK4 time-stepping of dR/dt = −R/S with the same update rule. (3) The cohort score against R, within four standard errors.
Not modelled. Sleep, interference, motivation, prior knowledge, question difficulty, feedback quality. This is a teaching illustration of two real effects, not a prediction for a real class and not advice about any individual student.
The classroom (floor, desks, board, chairs, students) was modelled and baked in Blender; rendering and all the maths run in Three.js r160.