Retrieval check answer. 5 chunks — one per hyphen-delimited group (f47ac10b, 58cc, 4372,
a567, 0e02b2c3d479) — down from 32 individual characters, the same move as lesson 2’s
FBI · CIA · NSA · IRS and lesson 15’s IPv4-by-octet: recode raw symbols into familiar units, and the
fixed slot count in working memory stops being the constraint.
Part 1 — the mapping
System 1 ≈ running a cached macro. System 2 ≈ the full planning loop. Fast, automatic, no visible deliberation vs. slow, effortful, step-by-step — the surface description fits cleanly.
Re-mapped onto this track’s actual machinery: running a matched macro is procedural firing (lesson 9) — a compiled IF–THEN disposition that executes the moment its trigger pattern is recognized, with no separate act of reasoning about whether to run it, exactly like the barista’s wand-angling habit or the chess player’s “IF pinned THEN don’t move it.” Falling back to the full planning loop is SOAR’s impasse-triggered universal subgoaling (lessons 11–12): no macro applies, or two apply with no basis to prefer one — that’s precisely SOAR’s definition of an impasse — and the response isn’t a different, ad hoc “slow mode” bolted on, it’s spinning up a genuine problem-space search whose entire job is to resolve the impasse. If that search succeeds, SOAR’s chunking would compile the result into a brand-new macro, so the identical situation never has to hit the slow path again — the agent’s macro library growing over time is chunking, not a separate “learning module.”
Part 2 — what dual-process theory hand-waves
Dual-process theory correctly names that processing shifts from fast to slow under some conditions — “when things get hard,” “when System 1 isn’t confident,” “when stakes are high” are the usual gestures. What it doesn’t specify is the actual triggering condition — a precise, checkable test for the exact moment control should hand off from one mode to the other. It’s a description of two styles of output, not a mechanism for the switch between them.
SOAR’s impasse is that mechanism, made explicit: the switch fires exactly when no production matches the current state, or more than one matches and nothing in memory prefers one over the rest, or an operator’s effects aren’t yet known. Not a vibe, not a confidence threshold tuned by feel — a structural test on the current match set.
The exact condition for the coding agent: fall back to the full planning loop precisely when either (a) no cached macro’s trigger pattern matches the incoming request closely enough, or (b) two or more macros match and nothing in the agent’s history or context prefers one over the other — a tie, in exactly SOAR’s sense. Anything short of that (a single macro matches cleanly) stays on the fast, procedural path; anything at or past it triggers the slow, deliberate one. That’s a condition you could actually implement and test against, which “when it feels hard” never was.
Why this matters beyond taxonomy. Dual-process theory is genuinely useful as a first-pass description — it correctly predicts that people (and agents built this way) will have a fast, error-prone-but-usually-right mode and a slow, effortful-but-more-reliable mode, and that the two produce systematically different failure signatures (System-1-style errors are misfires — a habit triggering on something that only looks like its usual pattern, exactly lesson 9’s pin-avoidance mislead; System-2-style errors are the well-documented human costs of effortful search under time pressure). What it can’t do is tell you, mechanistically, when the switch happens or why it happens then and not a moment earlier or later — and that’s exactly the gap a cognitive architecture is built to close, because an architecture has to actually run, which means the switch has to be a checkable condition on a data structure, not a description of the output it produces.
Where this goes: satisficing (lesson 19), representativeness (lesson 20), and this System-1/System-2 split are all accounts of judgment within a fixed way of seeing a problem. The concept still ahead — ecological rationality — asks a sharper question about all of them: a heuristic that looks like a bias in one environment can be the objectively correct move in another, which means “biased” was never a property of the heuristic alone. That’s next week’s opening move in Stage 5.