How each CorX Labs model was built and what it was measured on. Every number comes from the released files or from data the model never trained on, and where something has not been measured, the report says so.
A physics consensus that beats straight-line extrapolation and CLIPER at every lead from 6 to 120 hours, on 4,447 forecasts of 407 storms from seasons it never saw. It does not beat GraphCast or Pangu-Weather beyond 6 hours, and says so.
Three mixers over one residual stream: an exact window, a delta-rule state and a landmark-tree index. Attention pairs scale as N1.00, and recall ties full attention while the same model without its index falls to chance.
A rank-64 LoRA on 19 million tokens, merged into a 27B base: what it changed (voice, register, cultural grounding), what it did not (the base model's reasoning), and why factual use needs retrieval.
Pitch, timbre and breath in three streams that cannot share information until a late fusion trunk, with zero-shot cloning and a 32-step flow-matching decoder. No listening tests yet, and none claimed.
Architecture, tokenizer, three-stage training, and the memory arithmetic that fits three billion tokens through a single card.
Not peer reviewed. These are the lab's own technical reports, released under CC BY 4.0; the models and code are Apache-2.0. arXiv versions will be linked here when they are announced.
