Lessons 10–12 built up two architectures’ internal machinery in detail — activation formulas, subgoaling, chunking. It’s fair to ask: do these produce anything you can actually check against a real human sitting at a keyboard, not just an elegant story?
They do, and this is where cognitive architectures earn their keep as science rather than software design philosophy. ACT-R in particular has a long track record of fitting quantitative human data: reaction times in the low hundreds of milliseconds for simple retrievals, the shape of the power law of practice (performance improving fast at first, then ever more slowly — a prediction that falls directly out of the base-level activation math from lesson 10, not bolted on separately), and error rates in tasks like mental arithmetic or menu search, fit closely enough to be used in real human-computer-interaction design work (predicting how long a new UI layout will take users to learn, before it’s ever built). SOAR has matched timing and behavior in complex, well-structured task domains too — the classic case is fitting expert behavior in domains like airspace management and simulated piloting.
Where both architectures are consistently weaker: moments of genuine insight — restructuring a problem itself rather than searching harder within a fixed problem space or retrieving a stronger memory (the classic psychology example is the “aha” moment in insight puzzles, where the shape of the solution space suddenly changes rather than search within it succeeding). Both architectures are fundamentally built around searching or retrieving within a space of operators/chunks that’s already defined — modeling the moment a person invents a wholly new way to see the problem is a much harder fit for either.
Given that track record: which statement best describes their real, documented limitation?