Text to evidence
Notes
Workspace
The workspace now lives in the distillate repository under
content/lessons/text-to-evidence/, beside the other tracks and the learning
records. It once lived at docs/lessons/ in the Namaq repository, so that the
lessons sat beside the docs they cite. Links to Namaq files are now RepoLinks
to the repository on GitHub.
Preferences observed
- Learns by questioning the model, not the procedure. The session that started this workspace was a push-back — "but they hold a model value, right?" — and it was correct. Lessons should leave room for that rather than closing every question.
- Wants the rule's reason, and the repo agrees:
AGENTS.mdforbids comments that explain and sends the explanation to a document. Lessons are that document's cousin — cite the file or ADR, never assert from memory. - Reads Arabic; source excerpts can stay unglossed.
- Reads the text first and expects agents to. When a rule makes an agent stop on something a reader sees at once (who answers in a question and answer, whether two narrations are one event), the owner treats the rule as the defect. See learning record 0006 and ADR 0024.
Open threads for future lessons
- Should a heading be a citable passage?
PassageKindhasBODYandNOTE. A group heading in the Siyar's index carries a real model value (martyrdom at Badr) and no paragraph. Currently out of scope by consequence rather than by decision — that asymmetry is a design lesson. - Two authoring paths live at once. Three subjects (Abu Ubaydah, Talhah, al-Zubayr) come from the catalog alone; everyone else still comes from the seeds. Worth a lesson on the hand-off and how the catalog's authority differs before and after a seed entry is deleted.
- The read side.
sourceAccounts.tscaps pages per request at 9 and builds a section index by stripping the editor's bracketed headings. Lesson material on how the reader's constraints feed back into what a page may be.
Settled, no longer open
- Who owns a page — answered by ADR 0018
and taught in lesson 0002. The page store plan is
docs/plans/source-page-store.md; implemented in PRs #220-222. - Which invariants does this pipeline enforce, and which does it merely
offer? Taught in lesson 0003, using two instances the page-store migration
itself produced: an anchor-position check that counted paragraphs instead of
comparing the anchor's own number, and a dangling Prisma relation that
shipped past lint/tsc/1,701 tests and was only caught by a live query.
verifyExcerptsstill isn't wired intohistory:validate— that specific gap is unresolved, tracked as its own item below rather than folded into the lesson, since it's a workflow fix, not a design question.
Not yet fixed (not lesson material, just owed)
npm run history:verify-excerptsruns as a separate script; nothing calls it fromhistory:validateor CI, so a batch can pass validation with a citation excerpt that doesn't literally appear in its cited paragraph.