The University of Ottawa team won the EvaLatin 2026 named entity recognition (NER) competition for classical Latin texts, using prompt engineering on commercial models like gemini-2.5-pro and claude-sonnet-4-5. The performance was outstanding: first place in both the coarse-grained subtask (11 classes) and the fine-grained one (28 classes), with top scores across all evaluation regimes, strict and fuzzy.

The surprising bit isn’t just the academic victory but the method. Cross-lingual transfer learning was achieved without any fine-tuning, purely through prompt formulation. In essence, the researchers leveraged the linguistic capabilities already consolidated in models trained on dozens of modern languages, and redirected them toward Latin, a language underrepresented in digital resources. It demonstrates how advancements in the broader LLM community can power even the most niche tasks.

But there’s a hidden cost, and it goes beyond the token consumption. The LLMs used are commercial versions, accessible only via cloud APIs. Each invocation sends phrases from Cicero or Livy to third-party servers. If for a classical corpus this may seem like an academic footnote, the signal for those working under confidentiality constraints is clear: to use the best NLP technology, you hand control of your data to an external provider.

Sovereignty as the Missing Variable

The uOttawa experiment is a flawless technical demonstration, but it highlights a deliberate choice. All evaluation metrics focused on NER quality; none accounted for data residency, algorithmic transparency, or total cost of ownership (TCO) at scale. Yet for unpublished manuscript archives, for institutions operating under GDPR, or for culturally sensitive projects, self-hosted models become a strict requirement.

Who wins, then? Cloud providers, who see their centrality reinforced even in ultra-specific domains. Who loses? Resource-constrained labs that cannot afford rising API costs or on-premise models of comparable size. Moreover, the ability to audit linguistic biases introduced by the LLM vanishes: a closed model won’t tell whether certain systematic errors stem from a poor grasp of classical Latin or from a hidden flaw in the training data.

An alternative, of course, exists. Smaller, quantized models, hostable on local hardware, could be trained or fine-tuned for Latin NER with competitive results, provided there is investment in dataset curation and infrastructure. The trade-off is familiar: potentially lower performance, but absolute data control and predictable operational costs. The decision becomes strategic, not merely technical.

A Triumph That Questions Us

uOttawa’s top rank will stand in the records as an exemplar of transfer learning. But for those who look at LLM deployment not as a mere academic exercise but as an architectural choice, the lesson is twofold. On one hand, the immense potential of multilingual models even for dead languages. On the other, the fragility of an approach that entrusts cultural memory to external entities, with no guarantees of continuity, explainability, and compliance with information sovereignty. In the on-premise world, every Latin suffix processed locally becomes an act of independence.