DharmaOCR, a specialized optical character recognition model for Brazilian Portuguese, has demonstrated superior performance against newer, more capable generalist competitors, including Mistral OCR4 and Unlimited-OCR. Three months after its release and open-source publication, DharmaOCR's design effectively addresses persistent gaps in extraction quality and model stability for domain-specific tasks. Recent benchmarking confirms that domain specialization yields a measurable advantage even against recently released models backed by substantial research resources. In evaluations focused exclusively on Portuguese, DharmaOCR achieved a score of 0.925, significantly outperforming Mistral OCR4 at 0.798 and Unlimited-OCR at 0.7587. The performance gap stems from how model parameters are allocated. DharmaOCR concentrates its entire representational capacity on the vocabulary, syntax, and orthographic patterns of Brazilian Portuguese, whereas multilingual models must distribute parameters across a broader linguistic space. This structural focus allows DharmaOCR to correctly process culturally specific references and proper nouns that generalist models systematically misread. On complex documents such as Brazilian national examination essays, newer models failed to transcribe recognized names and phrases accurately, producing errors like transcribing Chico Buarque as Chico Barque. These failures highlight limitations in training data exposure rather than architectural deficits, as generalist models lack the concentrated optimization required for nuanced domain recognition. Beyond extraction accuracy, DharmaOCR addresses critical stability issues prevalent in generative OCR systems. The model employs a two-stage training pipeline: Supervised Fine-Tuning on diverse Portuguese sources followed by Direct Preference Optimization. While SFT aligns weights to domain specifics, DPO trains the model on complete output coherence rather than sequential token prediction. This approach mitigates text degeneration, a failure mode where models generate repetitive or incoherent text when faced with visually difficult inputs like small fonts or degraded scans. By penalizing incoherent outputs during training, DPO ensures DharmaOCR maintains reliability under production conditions, whereas generalist models frequently produce structurally unusable data when input ambiguity increases. Dharma's analysis underscores that while architectural advances raise the absolute performance ceiling for all systems, the structural logic of resource allocation remains unchanged. Specialization ensures that finite compute and parameter resources are directed toward maximizing performance in a target domain, a dynamic that generalist models cannot overcome regardless of their generation or capability. Dharma plans to continue integrating emerging architectures and training techniques while maintaining its focus on Brazilian Portuguese, reinforcing the principle that domain-specific optimization will remain a competitive imperative as OCR technology evolves.