wordsworth docs

wordsworth is a sovereign pipeline that turns large volumes of (mostly Dutch) government documents into a searchable, privacy-safe corpus. This docs/ tree follows the handbook docs contract; the project overview lives in the README, which this page points to rather than replaces.

Status: proof-of-concept, under construction. Design and delta specs live under openspec/.

Sections

• reference/ — facts: pipeline architecture, invariants, stack, and configuration.
• evaluation — the IR evaluation run: CLI, id-matching precondition (object_key = qrels doc ids), runtime needs.
• cli — the wordsworth API client: install, commands, ingesting a directory.
• explanation/ — reasoning and decisions.
• Key-user review 2026-10-28 — the console against four jobs: find topics, ask, judge relevance, share.
• NORA gap analysis — wordsworth tested against the "Anonimiseren & Pseudonimiseren bij de Bron" target architecture; verdict per requirement, build order.
• ADR-0005 — what wordsworth adopts, adapts and declines from that architecture.
• ADR-0007 — one token per type, also where a combination identifies: why a named quasi-identifier is reported and measured but never fused.
• meting 02 — 654 of 751 documents carry PERSON + DATE_TIME + LOCATION together, and what that number does and does not say.
• pilot report template — the published-dossier pilot's report, with two gate slots only Mark fills.
• access identity — logging in as a person, and why the readable header is never a source.
• index mapping — adding a field to the search index, and why a missing one fails silently rather than loudly.
• detection lists — allow/deny JSON lists, lists_hash in audit, the feedback endpoint.
• dataset pseudonymisation — CSV column path: profiles, per-attribute/per-record, NEN 7524-style output.
• pdf-redaction — the redacted/reveal PDF endpoints, the stale/value-not-located responses, how it is built.