Weight-tied depth recursion trades computation for effective depth without adding learned parameters. We pre-train 21 hybrid Gated DeltaNet (GDN) + grouped-query attention (GQA) transformers from scratch on the BabyLM 2026 Strict corpus under one tokenizer, data order, and optimizer, and use them to ask on which budget axis recursion actually pays. Under a matched data-exposure budget, recursion is a robust win on language-model loss: two exact parameter-matched pairs improve validation loss by 0.0252 and 0.0195 nats (means over three training seeds; seed sd 0.0015 and 0.0005), and a domain-balanced re-evaluation over the full 17.4M-token dev set puts the advantage at 0.0270 and 0.0196 nats, positive in all six dev domains under all three seeds (36 of 36 cells). Grammatical accuracy does not support a claim either way: replicated over the same three seeds the mean zero-shot BLiMP gain is +0.58 (sd 2.44) and +1.84 (sd 2.05) points, against a seed-noise floor of up to 2.80 points sd for a single architecture retrained—so the +3.31 and +4.21 obtained from one seed are not separable from training noise. Under a matched compute budget it is not: the same non-recursive twins trained to 1B tokens, using two thirds of the recursive models’ training FLOPs, beat them by 0.0117 and 0.0088 nats. Two further controls sharpen the mechanism. Untying the re-applied layers—same depth, same FLOPs, three times the parameters—recovers a further 0.0467 and 0.0399 nats, pricing what tying costs. Replacing effective-depth residual scaling with unique-depth scaling changes loss by only 0.0034 and 0.0003 nats, so the gain comes from tying rather than from the initialization the two differ in. A 25M-token checkpoint grid locates the recursive models’ crossover with their twins at 75–125M tokens in the 2:1 pair, before either model has completed one pass over the corpus, ruling out repeated examples as the mechanism there; the crossover moves later as effective depth grows, consistent with an early optimization cost. Recursion is therefore best understood as a data-efficiency mechanism for the fixed-corpus regime this benchmark defines, not a compute- or inference-efficiency one.