This is how NER should be done. I used Claude Code to add LLM integration. For the time being, I'm using GPT-5.6 Luna, which costs literally just pennies to process these requests.
Notice how this approach makes provenance visible and invites fellow historians to inspect the results.
