QENEX Law — Legal Knowledge-Base Acquisition & Curation Plan (v0.1)#

Purpose: collect, clean, and keep-current the most valuable legal-practice resources to power QENEX Law's drafting/review — built on properly-licensed material only. Scope: UK-first (England & Wales primary), internal-use corpus first. Status: planning; execution deferred behind Lab + unfinished work. Date: 2026-08-09

0. The one rule that governs everything: licence hygiene#

Every item enters the corpus only with a recorded licence + AI-use right. This is non-negotiable — an internal AI corpus built on copyrighted text you don't have rights to is a legal and reputational liability (ironic for a legal tool). Three buckets:

• Green (ingest freely): primary legislation, official government/regulator guidance under the Open Government Licence, and public-domain material.
• Amber (ingest only under a specific re-use/computational licence): court judgments via The National Archives Find Case Law (bulk/AI use needs their transactional/computational-analysis licence, not the default view licence).
• Red (reference, do NOT ingest into the shared corpus): commercial subscription content (Practical Law, LexisNexis, Westlaw), textbooks, and law-firm know-how. Licensed reading ≠ training/redistribution rights. Use for the human's own research, not the machine corpus.

1. Coverage map — prioritise QENEX's own document needs first#

Since it's internal-first, weight the areas your own paperwork actually uses:

2. Source tiers (UK, current & authoritative)#

Tier 1 — Primary law & official guidance (Green, OGL, free API/bulk):

• legislation.gov.uk — Acts & SIs, full API + bulk XML, revised + point-in-time.
• The National Archives "Find Case Law" — judgments, public API + bulk (Amber: get the computational/AI re-use licence).
• gov.uk regulator guidance — HMRC (incl. VCM manuals for SEIS/EIS), Companies House, IPO (patents/TM), ICO (data protection). OGL.

Tier 2 — Official model documents & standards (Green/Amber, check each licence):

• BVCA / UK Private Capital model investment documents — the standard for early-stage UK rounds; recently updated versions — high value, keep current.
• Companies House model articles & forms; IPO model/guidance; ICO template notices.

Tier 3 — Licensed commercial (Red — subscribe, don't ingest):

• Practical Law, LexisNexis, Westlaw — buy seats for human research; do not pipe into the corpus.
• Legal textbooks — purchase licensed copies; no corpus ingestion without rights.

Tier 4 — Secondary / open datasets (mixed licence):

• Open legal datasets (e.g. openlegaldata "awesome-legal-data"), academic open-access (SSRN, university repositories).
• Law-firm client notes / Lexology — monitor for change awareness only; don't wholesale ingest.

3. Acquisition methods#

• APIs / bulk downloads for Tier 1 (legislation, case law) — automated, repeatable.
• Structured fetch of official model docs (Tier 2) with source URL + version captured.
• Paid subscriptions / purchases for Tier 3 — logged as licences, kept out of the corpus.
• No scraping of copyrighted or ToS-restricted material. Ever.

4. Cleaning & normalisation pipeline#

1. Fetch (source URL, retrieval date, licence tag).
2. Extract text (OCR for scanned PDFs; keep original + text).
3. Segment into atomic units — clauses, sections, judgment paragraphs.
4. Tag metadata — jurisdiction, practice area, instrument type, enactment/decision date, source, licence, citation.
5. Deduplicate and link versions (supersession chains).
6. Extract citations / cross-references (legislation ↔ cases ↔ guidance).
7. Trust-score each unit (source authority + freshness) — this drives "draft from trusted templates, not free-form".
8. Chunk + embed for retrieval; keep human-readable source alongside every chunk.

5. Freshness — "newest" is a process, not a one-off#

• Monitor: legislation.gov.uk change feeds, HMRC/ICO/IPO manual updates, BVCA model-doc releases, new judgments.
• Scheduled refresh (reuse the existing scheduled-task pattern + qenex_websearch_*); re-score on update; flag superseded units.
• Each practice area carries a "last-verified" date; stale areas surface for review.

6. Provenance & storage (QENEX differentiator)#

• Anchor every ingested source + version through the dual-rail audit pipeline (BOSA RFC-3161 + OpenTimestamps) → provable "this is the text as of this date".
• Store: raw docs on IPFS (pinned); structured units + metadata in Supabase; curated templates/clause library version-controlled in Forgejo (qenex-law).

7. Tooling map (reuse existing QENEX infra)#

Acquisition → qenex_websearch_research/crawl/extract, qenex_http_call (allowlisted sources). Structuring → qenex_ai_proxy / qenex_multi_inference. Storage → IPFS + Supabase + Forgejo. Provenance → audit anchoring. Scheduling → scheduled tasks.

8. Phase plan#

• P0 (now, zero-cost): capture your own executed docs + the free Tier-1/Tier-2 sources you already rely on (legislation.gov.uk, HMRC VCM, Companies House, BVCA model docs) with licence tags. Start the clause library in Forgejo.
• P1: automate Tier-1 ingestion (legislation + case-law under proper licence) for the P1 practice areas; build the cleaning pipeline + trust scoring.
• P2: freshness monitoring + provenance anchoring; expand to P2 areas.
• P3: evaluate paid subscriptions for human research; revisit any external-product ambitions (with the licensing/UPL groundwork).

9. Immediate next actions#

1. Stand up the licence-tagging schema (green/amber/red) before ingesting anything.
2. Pull the free P1 sources (legislation.gov.uk, HMRC VCM SEIS/EIS, Companies House, BVCA model docs) and tag them.
3. Confirm The National Archives computational/AI re-use licence for case law before any bulk case ingestion.
4. Decide the Forgejo repo layout for the clause library vs. the raw corpus.