Thinking slow, by counting

Theory update. 2026-06-26 · win, with an honest negative · experiment AG

A count-native System 2: a metacognitive gate deploys a deliberate pass that overrides System 1 only when it's wrong, and does no harm when it's right. The gate is the win. The elaborate workspace is parked.

The question

Everything built before this is System 1. The char-order experts vote in parallel, a leader pops out, and the model ships it: fast, local, prepotent. That is the whole machine, and it has no second thought. Dual-process theory says a mind has another speed: a slow, serial, capacity-bounded workspace that engages only when the fast answer is suspect, decouples a few items from the stream, and overrides the reflex when the reflex is wrong.

We wanted to know whether a count model can have that second speed without becoming a different kind of model: no gradient, no backprop, no second training pass. And we wanted to judge it on the one test the working-memory literature insists on, the Engle signature: a real override improves accuracy where the prepotent answer is wrong and does no harm where there is no conflict. A System 2 that helps everywhere is not a System 2; it is just a better System 1. The discipline is that it must know when to stay out.

What we tried

Char next-char on text8, 12 million training characters, 300,000 held out. System 1 is a product-of-experts over short orders {2,3}: its vote activations the leaky accumulators, its argmax the prepotent leader. A dual trigger decides whether to think: deploy System 2 when calibrated confidence c = f·c (the NARS truth value from How sure is a count?) is low, or when Botvinick conflict is high (two strong answers fighting within System 1, or the fast leader and the wider context disagreeing). The wider {5,6}-order opinion is the top-down goal the fast voter under-weights.

When the gate fires, two operators were raced. The elaborate one is the faithful cognitive story: a capacity-4 focus, a serial race over leaky accumulators, inhibition-of-return so the loop advances, cognitive decoupling so the prepotent seed fades, a step budget, and a suppress-not-erase floor that commits the deliberate winner only if it beats the default. The trivial one is the cheapest possible deliberate act: consult the wider context and defer to it. Both share the same gate. Everything is read off counts in a single causal pass; the deliberate pass is a per-position loop over those same counts, no retraining.

The probe splits the stream. A conflict case is where the short-order leader and the long-order leader disagree, both contexts seen: the prepotent local bet fighting the broad context. On that subset, System 1's reflex is wrong 88.4% of the time. Exactly where a deliberate override should pay.

What happened: the Engle signature

The signature is textbook. On the conflict subset the override lifts accuracy by 0.38 (from 0.116 to 0.496, fixing the slice where the reflex is wrong nine times in ten) while leaving the no-conflict subset exactly untouched, 0.680 to 0.680, not a single case of harm. The gate produces the override-when-wrong, hands-off-when-right behaviour the working-memory work predicts. It fired on 71.8% of positions, and on the no-conflict cases it touched only the handful where even the wider context is wrong.

And the capacity bound is honest. Set the step budget to zero and the gated model emits the System-1 leader on every position: it equals System-1-only bit for bit, zero overrides. With no capacity to think, it degrades gracefully to the reflex. That is the suppress-not-erase guarantee made operational.

The honest negative

The two operators share the gate; they differ only in what they do once it fires. On the 31,290 positions where they disagree:

The workspace machinery does not earn its keep here. Where the elaborate race declines to switch or reaches for a third candidate, it is worse than simply deferring to the broader context: a third-candidate pick is wrong about 81% of the time. The mechanism that wins is the gate, the metacognitive decision to think, not the serial focus loop wrapped around it. So we park the workspace, not kill it. Its untested winning axis is the multi-step problem, where a single deferral cannot reach the answer and holding four items across serial cycles is the whole point: compositional reasoning, discourse coherence. Char next-token is a one-step decision. It cannot exercise a workspace.

The lesson

A count model can have a real System 2. The load-bearing piece is the dual trigger (calibrated confidence plus conflict) that decides when to think; once you decide, the cheapest deliberate operator, defer to the wider context, captures essentially all the win (+0.38 on conflict, ±0.000 on no-conflict, graceful to System 1 at zero budget). The elaborate serial workspace loses to the trivial deferral on this one-step probe and is parked for the multi-step tasks where its capacity is the point. The gate is the win; the workspace awaits its task.

The question the confidence work left open was what do we do when the fast answer isn't trustworthy? The answer, on a one-step decision, is the simplest default-interventionist one: notice the conflict, consult the wider context, override only if it disagrees. The metacognition matters more than the deliberation.

Lineage

Grew from how sure is a count?, whose NARS truth value is the metacognitive trigger that asks whether the fast answer is trustworthy; from one brain part, or many?, whose gate opens a route to a higher level; and from a vote that remembers, the leaky accumulators the deliberate pass runs over.

Thread: System 2, and the right combiner. The confidence post asked is the fast answer calibrated? This one asks and what do we do when it isn't? The ideas are Kahneman (default-interventionist), Stanovich (cognitive decoupling), Botvinick (conflict monitoring), and Engle (working-memory capacity as the override signal).