Solution: Why This Path Won't Let You Do 20 Lessons on One Topic in a Row

MCQ answer: (2) — Student B wins both, and the gap widens on the delayed test

Student B (spaced practice) scores higher on both the immediate and delayed test, and the advantage is typically larger on the delayed test — the spacing effect is among the most reliably replicated findings in the entire field of memory research, documented since Ebbinghaus’s original 1880s work and confirmed across hundreds of subsequent studies and multiple large meta-analyses. Option (0) states the intuitive-but-wrong belief most people actually hold. Option (1) is wrong specifically because total time spent is not the only variable that matters — its distribution matters independently. Option (3) inverts the actual, well-documented direction of the effect.

The reason the gap widens over time: both students’ memories decay after studying, but at different rates, for a mechanistic reason tied directly to Lesson 20’s deliberate-practice logic. Student A’s massed session lets later hours of study succeed via short-term familiarity — the material is still active in working memory from an hour ago, so “practice” during a crammed session partly consists of re-recognizing something still fresh, which (per Lesson 14) is a weaker learning signal than genuine retrieval. Student B’s spaced sessions, separated by real gaps, force each later session to be a genuine retrieval from longer-term storage (the material has actually left working memory in between) — the more effortful, more durable-trace-building act Lesson 14 identified as the active ingredient in retention. That stronger trace decays more slowly, so as both memories fade over the following month, B’s initially-more-durable trace stays legible far longer than A’s.

Part 2 — Why cramming feels productive despite underperforming

During a massed session, later study attempts genuinely do feel easy and successful — because the material is still sitting in working memory/short-term activation from minutes earlier, recognition is fast and fluent, and the student experiences a strong, immediate sense of “I’ve got this.” That fluency is real, but it’s diagnostic of short-term activation, not of a durably consolidated long-term memory trace — exactly the trap Lesson 20 named when it distinguished feeling-productive routine repetition from actually-effective deliberate practice. The subjective experience of ease during study is a notoriously poor predictor of long-term retention; if anything, the difficulty of spaced, effortful retrieval (which feels worse in the moment — more failures, more visible forgetting to confront) is the better predictor, precisely because struggle indicates genuine retrieval from longer-term storage rather than easy access to still-active short-term residue.

Part 3 — Mapping this path’s design to spacing and interleaving

Spacing: the routine’s core buffer policy — generating and delivering only a bounded number of unsolved lessons ahead of you (the roadmap’s “keep ≥8 unsolved lessons ahead, never more”), running weekly rather than dumping the full curriculum at once — is a direct spacing mechanism. It forces lessons on any given concept to be separated by real time (days to weeks, driven by your own solving pace) rather than consumed in one sitting, exactly matching Student B’s condition rather than Student A’s.

Interleaving: the roadmap’s decision to actively maintain four concurrent tracks (fp, bayes, cog, ml) rather than completing one track’s full spine before starting the next is a direct interleaving mechanism, at the level of entire domains rather than individual problem types within one domain (the more common lab-study version). Moving between a functional-programming puzzle, a Bayesian-statistics puzzle, and a cognitive-science puzzle in the same sitting is structurally the same move as Rohrer & Taylor’s interleaved math problem sets — mixed exposure instead of blocked completion.

A system optimized purely for immediate comprehension-per-lesson — how easy each individual lesson feels to follow right after finishing the previous one — would likely do the opposite of both choices: deliver an entire track’s lessons back-to-back (maximizing short-term context and continuity, minimizing the friction of re-orienting to a different domain each time) and dump many lessons at once rather than gating them behind a small buffer. That system would almost certainly feel smoother session-to-session. This lesson’s whole point is that “feels smoother in the moment” and “produces durable, transferable learning a month later” are not the same design target, and this path is explicitly optimized for the second one at some cost to the first.

Part 4 — A testable prediction from interleaving

Predicted short-term cost: when a bayes lesson immediately follows an fp lesson, initial engagement with the bayes material may feel slightly slower or more effortful than it would if you’d just finished several bayes lessons in a row and were still “warmed up” in that specific track’s vocabulary and mode of thinking — interleaving’s well-documented in-the-moment performance cost.

Predicted long-term benefit, and the test that would reveal it: months from now, faced with a novel problem that doesn’t announce which track’s technique applies (unlike a lesson explicitly labeled track: bayes), you should be comparatively better at correctly recognizing “this smells like a Bayesian updating problem” versus “this is really a work/span parallelism question” versus “this is a chunking/expertise question” — the exact skill Rohrer & Taylor’s interleaved math students showed a specific advantage in (correctly identifying which technique applies, not just executing a technique once told which one to use). A blocked, single-track-at-a-time version of this same curriculum would train you to execute each track’s techniques well within that track’s own context, but would give you comparatively little practice at the harder, more realistic skill of first figuring out which domain’s thinking a novel, unlabeled problem actually calls for.

The pattern

Rule: spacing and interleaving both trade a worse, more effortful in-practice experience for substantially better long-term retention and, for interleaving specifically, better ability to identify which technique a novel problem calls for — and both work through the same underlying mechanism as Lesson 14’s testing effect and Lesson 20’s deliberate practice: effortful retrieval under real forgetting builds a stronger trace than easy, fluent re-exposure.

Stage 4 closes here. You now have the full arc from Stage 0’s Marr’s-levels framing of minds as information processors, through Stage 1’s blackboard architectures, Stage 2’s integrated architectures (ACT-R, SOAR) and their real predictive track record, Stage 3’s memory and attention machinery, to this stage’s account of how skill and expertise are actually built and best trained — including, now, an explicit account of why this very puzzle path is built the way it is. Stage 5 turns to judgment and bounded rationality: why real cognitive agents, human or artificial, satisfice rather than optimize.