Whether Postponing One Grammatical Construction Delays the Others Is Decided by the Architecture

Children acquire grammatical constructions in a close to invariant order, and the classical explanation is cumulative complexity: a later-acquired construction contains an earlier-acquired one and something more. That account makes a prerequisite claim, and an observed order cannot test it, because the same order arises whether the earlier acquisition is required for the later or only simpler. Testing it needs the input to one construction postponed while the rest is held fixed, which cannot be done to a child. We build a learner in which it can. An artificial language of 242 word types is constructed so that each of its 7 constructions is the only place in the language where its own pattern of word classes and inflections occurs. Transformers of 0.63M to 10.74M parameters are trained from scratch on 200,000 sentences for 3,000 steps, at 10 seeds per condition, and a construction counts as acquired at the first of 70 checkpoints where it scores at least θ = 0.85 on 1,000 held-out minimal pairs and stays there. The intervention withholds one construction's sentences from the first fraction δ of the stream by exchanging them one for one with later filler, leaving the token budget, that construction's total occurrences and every other construction's arrival times unchanged; each run is paired with a reference run sharing its seed. Withholding a construction from the first half of training moves its own acquisition 1,384 steps later and the median other construction 9 steps, which is the instrument's own floor, measured on a construction that shares no manipulated feature. Those transfers are the same size whether the withholding covers a tenth of training or a half, while the effect on the withheld construction grows tenfold. A recurrent network of the same width and depth, on the same corpora, gives a dense matrix instead, and its transfers scale with the postponement rather than staying fixed, rising from 45 to 729 steps over the same range of δ. Whether a postponed construction takes the rest of the grammar with it is a property of the learner, not of the language.