Accurate date normalisation—mapping partial or abbreviated date expressions to canonical era–year–month–day forms—is a prerequisite for reliable `<date>` TEI encoding, and a task where LLMs offer evident potential in sparse-data historical settings. Yet systematic evidence of where automated inference succeeds and where it requires human intervention remains absent, hampering principled deployment in scholarly editing. This challenge corresponds to D4—semantic recognition—in the TEI encoding quality framework of Strutz (2026), the dimension least amenable to automation.
We demonstrate a controlled experiment using the Shōsōin Document—an eighth-century Japanese administrative corpus whose interleaved records from multiple concurrent scribal projects pose particular difficulties for contextual reasoning—as a testbed for LLM-assisted date normalisation. Our central methodological contribution is that context must be aligned to scribal work units rather than physical page boundaries: splitting context at page breaks severs the era and year information necessary to resolve partial date expressions. Where even a scribal unit lacks era or year information, we inject an anchor passage from the immediately adjacent unit. A 100-instance gold standard spanning four date pattern types was evaluated using Claude Sonnet 4.6, guided by a structured annotation guideline as the sole task specification—without fine-tuning.
This architecture achieves 91.8% overall accuracy (90/98 evaluable instances): 100% on complete and abbreviated forms, 95.5% on year-omitted expressions, and 78.1% on year-month-omitted expressions. We introduce ΔC, a distance metric quantifying how far beyond a date's own scribal unit the necessary contextual evidence resides, and show that anchor injection effectively resolves ΔC = 1 cases. Residual failures are structurally irreducible: month ambiguity in multi-month ledgers, where the same day number appears in consecutive months, cannot be resolved from textual context alone. This failure taxonomy supports a principled human-in-the-loop design—the ΔC metric flags cases for expert review, enabling automated processing of the majority while concentrating human effort where it is genuinely indispensable.