When a habit becomes a thought

Theory update. 2026-06-26 · qualified win · experiment AH

Stability, not error, promotes a mastered construction into an explicit, slot-addressable concept that answers questions the flat count cannot. The cost is a transient dip in accuracy, exactly the U-shaped curve Karmiloff-Smith predicts.

The question

Everything this cortex has built is implicit. A count or a cluster maps input to output, and its parts are not separately addressable. You can ask the construction what comes after oxford? but you cannot ask it show me every construction that fills a UNIVERSITY slot, regardless of its verb: the slot is not a thing you can name, it is a side effect of a count.

Karmiloff-Smith's Beyond Modularity says a mind does not leave knowledge in that state. Once a behaviour works implicitly, it gets spontaneously re-described into a more explicit format whose parts ARE addressable and recombinable. And the trigger is not error; it is stability. A behaviour that has stopped changing is "done," and the system re-represents it so a slower process can bind, compare, and recombine its parts. We asked whether that mechanism is expressible in a count world, and whether it buys a capability the implicit count cannot have. It mattered because System 2 needs operands: explicit things to deliberate over. Redescription is where they come from.

What we tried

text8, 16 MB, 2.73 million words. The top 10,000 words each get an online category from predicting the kind's signatures and leader clustering. The implicit grammar holds 10,000 frames, 5,380 of them open-slot candidates. Beside each frame we put a stability monitor: a count-native mastery detector that watches whether the leader category and the next-token distribution have stopped moving over the last several exposures (same leader, drift below a floor, for N exposures running). No target, no error term. Mastery by settling.

On stability, not error, the frozen co-firing pattern is promoted into an explicit slot-object: a named node with a separately-addressable role (the frame word) and filler (the slot category, its member words, and their probabilities). A registry holds these and, crucially, an inverted index from slot-category to frames that the frame-keyed count never exposes. Promotion copies the stable counts; it does not retrain them. We streamed all 2.07 million open-slot exposures once and 5,217 constructions were promoted. One pass, fixed seed, about a hundred seconds on CPU.

What happened

The explicit registry answers three queries that are structurally impossible on the flat frame-to-filler table:

inverted slot lookup   slot 347 {university, college, oxford, cambridge, professor}
                       <- "oxford ___", "harvard ___", "yale ___"
role substitution      "oxford ___" --swap role--> "harvard ___"   same slot 347
slot analogy           a : b :: c : ?   resolved by binding the shared slot

Twelve slots are shared by two or more promoted constructions; ten are genuine content slots, giving 17 cross-frame substitution pairs over real categories: university↔college, example↔instance, semitic↔semitism. The one function-word mega-cluster is the known leader-clustering artefact, and it is the honest noise floor, not the result. Answering any of these on the flat table requires rescanning every frame and re-deriving its leader at query time, reconstructing the explicit layer to answer the query. The parts are not addressable. That is exactly the gap, and the redescription closes it.

The second signature is the dip. When prediction hands over from the smooth implicit count to the freshly-promoted explicit form (which first predicts through the slot type alone, discarding the construction's specifics) accuracy should regress, then recover as the explicit object re-binds its own dominant filler.

A textbook U: 0.155 → trough 0.051 → recovered 0.181, ending above the smooth implicit baseline. The handover to the explicit slot-type costs about ten accuracy points, holds low while the form is purely compositional, then recovers once the explicit object re-binds its specific filler, and the recovered form is both manipulable and no worse a predictor than the count it grew from.

The honest caveat

The recovery is a consequence of the modelled mechanism, not an emergent surprise. We built the explicit form to first discard the construction's specifics and then re-bind them; a purely frozen snapshot with no re-binding would dip forever. So the finding is not that a U falls out of nowhere; it is that the Karmiloff-Smith mechanism is expressible and self-consistent on counts, with manipulability the real win. The dip is the predicted signature of redescription-then-integration, and the recovery is its claim: an explicit concept can be both nameable and as good a predictor as the habit it came from. Judged on raw accuracy during its dip, the explicit form loses; judged on whether it lets a slower process ask a question the count cannot, it wins outright.

The lesson

Put a stability monitor beside every count (mastery is the leader and the distribution settling, no target, no error) and on mastery, redescribe the stable construction into an explicit slot-object in a registry with an inverted slot index. You gain a query surface the flat count structurally cannot offer: what fills this slot anywhere, substitute the role, complete the analogy, for free, online, no gradient. Expect a brief accuracy dip at each promotion; it recovers, and can exceed the count, once the explicit form re-binds its specifics.

This is how an implicit habit becomes an explicit thought. And it is the missing half of the System-2 story: the deliberate pass needs manipulable operands to deliberate over, and redescription is what mints them.

Lineage

Grew from words that lower the cost of letters, the first counted concept that earned its keep; predicting the kind, not the word, whose online categories the slots route through; and grammar is just counting, made productive, the implicit constructions this promotes into explicit slot-objects.

Thread: representations. A redescribed construction is the explicit, manipulable operand the serial workspace needs, the implicit-to-explicit, count-to-concept bridge. The idea is Karmiloff-Smith's representational redescription.