Text to evidence

Reason for learning the source-extraction pipeline: to evolve its design.

Not to author batches by hand, and not to review an agent's output line by line. The goal is to reason about the model itself — what a claim is allowed to back, where the seam between a work's text and Namaq's structure sits, which parts of the pipeline are load-bearing and which are conventions that could change.

What this means for lessons

• Teach seams, not procedures. npm run history:validate is a fact to know; why the validator checks paragraph counts is the lesson.
• Every design claim cites the file or ADR that holds it, because a design argument that stands on nothing is the thing this repo already rules out for historical values.
• The success test is predictive: given a proposed change to the model, say what it would break and which rule already answers it.

Origin

Grew out of a question the pipeline did not obviously answer: the Siyar's group headings (شهداء بدر, السابقون الأولون) look like they hold a model value — that someone was martyred at Badr — yet they have no prose to extract and no subject to attach to. That is an eligibility question about the model, not an extraction question, which is what set the mission here.