Training data for modern language models is typically bucketed by coarse domain labels or filtered by quality heuristics, with no principled fine-grained taxonomy of conceptual or affective states for coverage. We propose the Odu-256 Training Curriculum: a 256-state product taxonomy derived from a 16-by-16 archetype product space. Each state is identified by a 4-bit-by-4-bit code; the 16 diagonal cases are resonance states (stable single-archetype), and the 240 off-diagonal cases are transition states. We describe how to use this taxonomy for stratified training (uniform-over-state or corpus-weighted sampling), stratified evaluation (per-state metrics that reveal which conceptual states the model handles well or poorly), and curriculum ordering. The taxonomy aligns by construction with the Vortex-keyed routing architecture (companion preprint): the 16 archetypes are the routing primitives, so per-Odu evaluation yields per-expert evaluation. We outline a pretraining and evaluation program comparing Odu-stratified mixtures against uniform random sampling on standard benchmarks and proposed diversity metrics. Empirical results are deferred to v1.1.
Weslyn Cory Whitehead (Wed,) studied this question.