Correct: they predict well-structured, well-practiced task performance fairly precisely, but struggle with genuinely novel insight or creative restructuring.
The other options overstate in the wrong direction: ACT-R’s single strongest, most-replicated success is reaction-time and error-rate prediction on exactly the well-practiced tasks option 1 claims it fails at — the power law of practice is one of the most robust quantitative fits in cognitive science. Option 3 is also false — ACT-R and SOAR both model mental/cognitive tasks (retrieval, arithmetic, decision-making) as centrally as, or more than, motor tasks. And far from abandoned, ACT-R specifically remains in active use today, including in applied HCI work predicting real interface learning curves before a design ships.
Why the insight weakness isn’t incidental — it follows directly from Stage 1–2’s own architecture. Both ACT-R and SOAR are built around search or retrieval within a fixed, already-defined space: SOAR searches within operators available in a problem space (lesson 11); ACT-R retrieves among existing chunks weighted by activation (lesson 10). A classic human “aha” moment — the nine-dot puzzle solved by drawing lines outside the assumed grid, or a joke’s punchline reframing everything before it — isn’t “searched harder within the space,” it’s the space itself being restructured: the constraint that was implicitly assumed (stay inside the dots) turns out not to be a real constraint at all. Neither architecture has a built-in mechanism for questioning whether its own problem-space definition — the very thing search happens inside — is the right one. That’s not a bug either team failed to notice; it’s a direct, structural consequence of committing to “intelligence is search/retrieval within a space” as the foundational bet, the same bet that buys the precise, checkable predictions on well-structured tasks in the first place. The tradeoff is real, not a coincidence.
Why this matters as evidence, not just design taste. This is what separates a cognitive architecture from an arbitrary agent design: it makes numerical, falsifiable claims (a specific reaction time, a specific learning curve shape) that can be — and have been — checked against real humans and found to hold in a well-defined range of tasks, and to fail in a different, characterizable range. That falsifiable-and-partially-right status is exactly what “unified theory of cognition” was supposed to mean for Newell, and exactly why these architectures are still argued over rather than dismissed as untestable philosophy.
In a modern LLM agent harness, this same asymmetry shows up directly: an agent given a well-defined tool space and a clear goal often performs reliably and predictably (the harness’s “search,” across tool calls, is well-structured) — but an agent asked to notice that the task itself is framed wrong, or that a completely different approach would dissolve the problem rather than solve it, is exactly the harness-design equivalent of the insight gap this lesson describes.
Where this goes: Stage 2 closes here, having built two full architectures from the ground up and checked their track record against real humans. Stage 3 turns to the psychology directly — memory and attention as studied independent of any one architecture’s specific claims about them.