Finalize humatheque extraction prompt and task constants

[gegedenice]

No description provided.

[davanstrien]

Hi @gegedenice thanks for this, and sorry it sat so long!

The benchmark core has moved quite a bit since April (a few merged PRs touching run.py/cli.py), so this branch now conflicts and wouldn't apply cleanly as-is.

The bigger qs is that the PR is closely tailored to the Humathèque thesis task whereas ocr-bench core tries to stay task-general. So I don't think it should merge into the core modules in this shape.

IMO there's a really good idea in here I'd love to keep: the deterministic, reference-based scorer (exact / list / fuzzy fields) as an alternative to the LLM-as-judge path. That generalises well and it's something the benchmark doesn't have yet.

Two ways forward, whichever suits you:

• a smaller, config-driven version where the task specifics (dataset, fields, vocab) live in config rather than core, plus a short note on what it does and how you checked it; or
• keep the Humathèque setup as a worked example/recipe rather than in core.

let me know what you think