The reason behind every rule in a session

An error is not a neutral data point. Terrace compared four ways of introducing the wrong option and found that introducing it early and faint produced between 2 and 9 errors during acquisition, while introducing it late and at full strength produced between 175 and 228. The count was not the interesting part. The subjects that learned through errors kept producing bursts of them long after the discrimination looked settled, responded faster but less reliably, and reacted to the incorrect stimulus as though it had become unpleasant. Errors leave a residue — in what gets learned, and in how the learner feels about sitting down to the task tomorrow.

Terrace never claimed that error-free teaching is always the better trade, and neither do we. Reviewing 46 experiments, Cengher and colleagues found that the class of help Interlaza uses — built into the display itself rather than added on top of it — was more effective than the alternatives in 14 of the 15 studies that compared the two. But they also found that a procedure which tolerates a couple of errors won its own comparisons. Minimising errors is one measure of efficiency among several, and which one matters depends on the learner in front of you. It is our default, not a law.

Carroll, Kodak and Adolf held everything constant except the delay between a correct response and its consequence. Ten seconds was enough to do damage: one of the two children lost efficiency, the other stopped acquiring the targets altogether and had to be re-taught with immediate reinforcement. Their reading is that the gap lets whatever the child happens to do inside it get reinforced instead of the response you meant.

Immediate praise partly bridged the delay for the child who already had a long history of praise being followed by something good — which is a reason to pair the app's celebration with something real in the first few sessions, not a reason to relax about the delay. The same caution shows up from a different direction: delayed feedback interferes with derived relations emerging at all, and derived relations are the part of learning this platform exists for.

With two options, a child who simply always picks the card on the right is correct half the time — and that half gets reinforced. The score climbs while nothing at all is being discriminated. Grow and LeBlanc put the arithmetic plainly: two comparisons reinforce a side preference around 50% of the time, four bring it to around 25%. Green recommends at least three for most teaching purposes.

Sidman goes further, and for a different reason. With only two options you cannot tell whether the child selected the correct one or rejected the other, which makes the result ambiguous as data even when it is high — he asked for experiments to be repeated with three or more comparisons for exactly this reason. On a trial that is meant to measure rather than teach, that is not a difficulty setting. It is the difference between a reading and a shrug.

A stage is a few dozen trials. Uniform randomness balances out in the long run, and a few dozen trials is not the long run — chance alone will happily put the answer on the right four times running, which is all it takes for "always the right one" to start paying off. Green describes the failure directly: when the correct option turns up more often in one place, the choice comes under the control of the place instead of the sample, and the child can look accurate without ever having looked at what they were matching.

One honest qualifier. Leaf and colleagues compared three ways of rotating stimuli and none of them produced faster acquisition than the others. Counterbalancing is not a speed trick, and we do not sell it as one. Its value is that it removes a competing explanation for a good score — so when the score is good, it means what it says.

Teaching one example teaches that example. Stokes and Baer's review of the generalization literature found that the way to get a behaviour to appear outside the exact training conditions is to train enough different instances that the induction forms — and, encouragingly, that a fairly small number is often enough. Rehfeldt and colleagues state the same idea as a procedure rather than a finding: a relational skill counts as established only when the learner performs on a set that was never directly trained.

Two things to be straight about. Interlaza rotates exemplars within the trained set; the stronger version in the literature — train one set, test a set that has never been trained, and if it fails, teach that one and test another — is a policy we have not built. And a number of concepts currently carry two instances, which is the minimum the literature reports as often sufficient rather than a comfortable margin.

A probe is a measurement, and reinforcing a measurement changes the thing being measured. Sidman and Tailby's central argument is that a child performing well under training tells you the training arrangement is working, and nothing more than that; whether the relation holds once the support is removed is a separate question that needs a separate, unreinforced test. Expect the score to fall when the support comes off. That fall is the reading — it reports how much of the earlier accuracy was resting on the arrangement.

There is a cost to how we do this, and we would rather name it. In Sidman's own procedure probes were interleaved into a still-reinforced baseline, with the reinforcement thinned in steps beforehand, precisely so that a probe could not be told apart from an ordinary trial. Interlaza runs probes as their own block, which is a more noticeable arrangement than his.

Stokes and Baer read some 270 studies and found that close to half the field's approach amounted to training, hoping, and reporting generalization when it happened to appear. Their conclusion is the one we build on: treat generalization as something that never comes free, and measure it deliberately — including when it does not appear. "Mastered in training, fails with a new picture" is more useful to an Instructor than a mastery score on its own, because it names the next thing to do.

Deliberately absent for now: Interlaza measures generalization but does not yet teach it. New instances only ever show up in unreinforced probes, so there is no trial in which generalizing is itself reinforced. Stokes and Baer's own list asks for both halves, and we have one.

Generalization across time is what maintenance means, and it is one of the conditions Stokes and Baer insist has to be programmed rather than assumed. A criterion met on a Tuesday is a statement about Tuesday. Widening the gaps also makes practice progressively less predictable, and that unpredictability is part of the point: it is the property that lets learning survive outside the schedule that produced it.

This is the reason the platform exists. Sidman found that teaching a handful of conditional discriminations produces a far larger number of relations the learner was never taught: in the 1982 study, nine taught relations yielded twenty-seven emergent ones. Relational Frame Theory extends the same finding past sameness to other patterns of relating — opposition, comparison, part and whole. What is being taught in that case is not a set of items but a way of responding, which is exactly why the tests that matter are the ones over material the child has never practised.

The two traditions are separate evidence bases and we do not add them together into one impressive number. Dymond and colleagues counted the relational-frame literature and found 62 empirical papers between 1991 and 2008, 42 of them addressing the central claims, with the bulk of the data collected from adult university students rather than from young learners. Sameness has dense support; opposition has some; comparison less. The further frames the engine can generate — hierarchy, time, space, cause, perspective — are our extension by analogy, not established practice, and we would rather say so here than have you discover it later.