YvyrAI: A Spanish-First Recurrent-Deliberation Language Model Architecture with Internal Verification, Conditional Repair, and Adaptive Compute

We present YvyrAI, a Spanish-first decoder-only language model architecture centered on a recurrent module called the DeliberationBlock. Rather than treating scale only as parameter count, YvyrAI explores a second axis of scaling: adaptive internal computation. The DeliberationBlock iterates a shared Transformer block within a single forward pass, using a learned controller to maintain a gated planning state, access a compressed latent scratchpad, estimate an internal verification signal, apply conditional repair updates, and decide how many refinement steps to perform before emitting the final hidden representation. Around this core, YvyrAI integrates modern language-model components including Grouped-Query Attention, RoPE positional embeddings, SwiGLU activations, RMSNorm, optional Mixture-of-Depths routing, sliding-window attention, low-precision training support, KV-cache compression, a Spanish-first data pipeline, a multi-term training objective, and an implemented Spanish evaluation harness. The contribution is architectural and methodological. YvyrAI is not presented as a trained high-capability model, nor does this paper claim superiority over existing systems. Preliminary experiments on a 15M-parameter configuration were used only as temporary engineering validation to verify that the tokenizer path, recurrent loop, auxiliary losses, and evaluation harness execute end-to-end. Large-scale benchmarking is deferred until a sufficiently larger training run is available. This preprint documents the architecture, implementation strategy, training methodology, evaluation framework, limitations, and scaling roadmap for the YvyrAI Research Project in Paraguay.