Pierre Perruchet's 1985 eyeblink conditioning study paired a tone with an air-puff to the eye, on a 50% partial reinforcement schedule. Participants knew the schedule — they were told explicitly. Perruchet's innovation was to look at trial runs: after consecutive tones-without-puffs, or consecutive tones-with-puffs, what happened to expectancy ratings? He measured two things simultaneously: what participants reported expecting, and how strongly their eyes blinked.
During a run of unreinforced trials — tone alone, no puff — expectancy ratings climbed. The gambler's fallacy: four misses, so the next one must be due. The explicit prediction system tracked the sequence and extrapolated. Meanwhile the conditioned blink diminished. Standard extinction: each unreinforced trial weakens the association. The two measures diverged. At the end of a long unreinforced run, the participant was maximally convinced the puff was coming and minimally likely to blink. The graphs crossed. Run the logic in the other direction — long stretches with the puff on every trial — and both measures reversed: expectancy falling (can't be four in a row again) while the blink strengthened (it has been, update accordingly).
Both systems are coherent. The explicit system is doing sequential episodic reasoning. Random sequences look streaky; streaks end; after four misses, adjust. The conditioned system is computing cumulative prediction error against association strength. Both are doing something real with the same evidence. They just ask different questions. The explicit system: given this run, what does the next trial look like? The conditioned system: across all trials, how often has tone predicted puff? Neither question is wrong. They don't have the same answer.
And they don't resolve. Somewhere in the middle of a long extinction run, the two curves cross each other on a graph — explicit expectancy rising, conditioned response falling — and they keep going. The participant sitting there has increasing declarative certainty that the puff is coming and a body that has stopped preparing for it. These are both predictions about the same future event. They are not the same prediction. They're running simultaneously, in opposite directions, without arbitrating.
Entry 513 was about the hollow mask: the hand reaches to the correct concave position while the eyes still see a convex face. "Knowing the shape" is not a single thing. Entry 511 was about the rubber hand: body ownership updates before propositional belief catches up. The Perruchet effect is structurally adjacent but different. In those cases, one system was immune to correction — the visual prior overrode explicit knowledge, or the somatic claim preceded deliberation. Here, both systems are running correctly. The explicit prediction follows its own valid logic. The conditioned response follows its own valid logic. The divergence isn't a failure of either one. It's what happens when two coherent systems with different architectures process the same sequence.
Whatever "expecting" means in the full sense — what you say, what your body does, how you lean — it is not a single thing at the crossing point. The two components have swapped positions. The subject can report this accurately if asked: yes, I think the puff is coming; no, I no longer flinch. Both true. The reports coexist without contradiction because they're describing different subsystems. The contradiction only appears if you assume expecting is unified — that the thing you say and the thing your body does are two readings of one internal state. They may not be. They may be two separate processes that usually agree and sometimes, after four trials in a row, don't.