Curriculum Order and the Geometry of Learned Representations: The Basque First Hypothesis

The field of large language model development has optimized aggressively along two axes: scale of compute and volume of training data. This paper proposes a third axis, almost entirely unexplored at frontier model scale: training language curriculum order. We argue that English-first pretraining — the de facto standard — builds reasoning capacity on a morphologically impoverished foundation, encoding argument structure through positional heuristics rather than explicit grammatical representation. We introduce the Basque First hypothesis as a proof of concept: that beginning pretraining with typologically distant, morphologically rich languages forces the construction of richer internal representations of predicate-argument structure before any English token is encountered. We then extend this argument to its logical conclusion — a purpose-built synthetic training language designed to encode logical structure, evidentiality, quantifier scope, and predicate-argument relations without the historical accidents and ambiguities of natural language. This synthetic substrate, invisible in the model’s outputs but structural in its weights, functions as what we term the dark energy of the model: inferred not from observation but from the coherence and reasoning quality it produces. Recent interpretability work — Anthropic’s finding that verbalizable representations form a global workspace (the J-space) within language models — supplies both a coordinate system for this claim, since the substrate’s influence is predicted to reside in the non-verbalizable complement of that workspace, and a measurement instrument for testing it. Finally, we identify an emergent flywheel: AI systems trained on such substrates will generate formal proofs, structured arguments, and logical derivations that become the highest-quality training corpora for the next generation of models — compounding reasoning quality independent of raw scale.