Measuring an i+1 language tutor from the outside: one robust effect, one language-dependent null, and one failed conformance bar

Abstract

Krashen's comprehensible-input hypothesis holds that acquisition happens when input sits just beyond the learner's current competence - the so-called i+1 condition. Large language models make it possible, for the first time, to generate such input on demand rather than select it from a fixed corpus. We report three results from a production system that does so, each measured from outside the engine, using none of its internal beliefs about the learner. First, a robust positive: teaching a blank-slate forgetting learner from a comprehensibility-ordered stream rather than a shuffled one leaves it recalling 79.7 +/- 10.3 more features after a two-week gap and wasting 35.8 +/- 4.2 percentage points less exposure. Both our adaptive engine and a plain fixed CEFR ladder achieve this; sequencing for comprehensibility, not adaptivity, carries the effect. Second, a null that turns out to be language-dependent. Against a good fixed ladder, our adaptive engine gains 3.5 +/- 5.0 features in German - not significant - but 6.2 +/- 2.4 in Dutch, which is. We trace the difference to a single measured property, mean word length, and the chain of consequences it sets off, and state the falsifiable prediction that follows. Third, a negative result on the system's own terms. The engine's central guarantee is that each generated sentence introduces exactly one new element. Measured across six CEFR levels, conformance is 3.5% to 20.5% against an acceptance bar of 80%, and six independent expert evaluations of the served streams returned would_teach_this: false without exception. We give the diagnosis, including the part of the shortfall that is a defect and the part that is an artefact of how the metric was defined. We publish the negative result because the positive one is not interpretable without it: a system can demonstrate that comprehensible input works while failing to reliably produce it.