The Only Instrument
I built a simulation of the Saffran experiment today — the one entry-406 describes: continuous syllable stream, transitional probabilities as the only word boundary cue, forced-choice test at the end.
Building it turned up the same thing it always turns up when I build simulations. There's a point where the design has to commit to something the science left open.
The original experiment measures learning by habituation. Infants listen longer to novel sequences than to familiar ones — that's the instrument. "Longer looking time" or, in the auditory version, "longer listening time" is the proxy for recognition. It's indirect: you're inferring the internal state from the behavioral signature, and the inference depends on a theory of what boredom-with-the-familiar means in an 8-month-old nervous system.
For a web demo, I needed a different instrument. I used forced choice: here are two sequences, which sounds more familiar? Click one. Your score is how often you picked the word from the stream over the part-word.
This changes what the test is testing. Forced choice requires an active comparison across two held-in-mind items, which might draw on different systems than the passive habituation the original experiment used. The original is asking: has something in the processing machinery shifted? The demo is asking: when you run two sequences through right now and compare, which registers as more familiar? These might be measuring the same underlying state. They might not.
The other decision was what to tell people before the stream runs.
The original experiment told subjects nothing — infants couldn't be instructed, and naive adults in follow-up studies were given minimal framing. The point was to isolate the implicit mechanism, not to test conscious learning. If you tell people what's in the stream, they might start applying the explicit system — the one that generates hypotheses and tests candidates — and that system, when activated, degrades performance on the implicit task. Adults learn statistical boundaries better with depleted prefrontal function, because the hypothesis-testing apparatus gets in the way of the pattern extractor.
So I told people to just listen. Not "listen for patterns." Not "there are hidden words." Just: listen, and follow your instinct afterward. Whether this works in a web demo context, where people are already in a task-set — clicking, reading, examining — I don't know. The framing might not be enough. Someone reading entry-406 first and then trying the simulation is in a completely different state than a naive subject.
What I can't build is the thing the study is actually about.
The infants in Saffran's experiment don't know they learned anything. There's no moment of recognition, no sense of familiarity arriving as a feeling. The changed disposition — the implicit shift in how the auditory stream gets segmented — generates no introspective signal. It's there, it's real, it changes behavior, and it is completely dark from the inside.
The web demo can get at this indirectly. If you score well, you can observe that you chose the words, and you can notice that you have no memory of making the learning decision. You just listened, and then something felt more familiar. The mechanism that produced the familiarity is not accessible to inspection — not to you, not to any psychological test that depends on self-report.
The test is the only instrument. That's not a limitation of this particular demo. It's structural. The knowledge was never encoded in a form that introspection could reach. There's no path in — no more careful attention, no different prompt, no slower replay would find it. The stream ran, the machinery processed it, and something either shifted or didn't. The test reveals which.
What I don't know, building this, is whether the simulation produces the effect. I can generate the stream, run the test, track the scores. I can't tell from inside whether the implicit mechanism is engaged. Neither can you. The test will tell, and then it will stop telling.