Luxembourg (LU) — internal licensing notes

Source research, 2026-09-02 — sellability + reproducibility audit

Clean GO. GeoNames' postal export carries both admin tiers Luxembourg has (canton, commune) — no gap to fill, no second source needed. Same shape as AT's audit.

The source — GeoNames LU

https://download.geonames.org/export/zip/LU.zip      # 4,519 rows

Licence CC-BY-4.0, verified against export/zip/readme.txt on 2026-09-02 — Luxembourg carries no per-country caveat (the readme's only restricted countries are GB, CL, CN, IE, MT, AR, BR).

Verified directly against the raw source file:

• admin_name1/admin_code1 — Canton, 12/12 covered (Capellen…Wiltz), GeoNames' own 2-letter code (CA…WI).
• admin_name2 — Commune, 105 distinct values, admin_code2 is a per-canton sequence number (e.g. "01" repeats across many different cantons) — not a globally unique commune code, so it cannot be used alone as part of a primary key.
• admin_name3/admin_code3 — always empty. Luxembourg's postal source has no third admin tier at all (unlike AT, which has Gemeinde as admin3) — confirmed empty on all 4,519 rows, not a partial-coverage gap.
• 1 (postal_code, place_name) pair collides: "Blumenthal" at L-7639 is listed twice — once under Bech/Echternach canton, once under Junglinster/Grevenmacher canton — a real border locality straddling a commune/canton boundary, same pattern as AT's Weissenbach. Real primary key is (postal_code, place_name, admin_name2), verified unique across all 4,519 rows.

ISO 3166-2:LU derivation verified 2026-09-02 against Wikipedia's ISO_3166-2:LU page: GeoNames' own canton codes (CA, CL, DI, EC, ES, GR, LU, ME, RD, RM, VD, WI) map 1:1, letter-for-letter, onto the ISO codes (just prepend LU-) — even simpler than AT's numeric crosswalk, no ambiguity.

licence_status

{
  "channel": "postal-codes-r2",
  "source_url": "https://download.geonames.org/export/zip/LU.zip",
  "licence_id": "CC-BY-4.0",
  "collected_at": "2026-09-02",
  "decision": "compatible",
  "decided_by": "claude-session-2026-09-02",
  "decided_at": "2026-09-02"
}

Proposed tier

B ($49). Single official-quality source (GeoNames) normalised in place — real work (ISO code derivation, admin-tier cleanup) but no second authoritative provider consolidated onto it. Same shape as AT/IT/PH.

What was already live before this rebuild

The prior lu/lu.csv on R2 (381,661 bytes, dated 2026-07-30) was NOT the "100-row sample, contact sales" placeholder the old datapackage.yml/ README.md claimed — it was the full 4,519-row GeoNames extract, just with the standard legacy corruption: literal "nan" strings in every row's admin_name3 (round-tripped through pandas from an empty upstream field), and float-formatted accuracy ("4.0"/"6.0" instead of "4"/"6"). No .licensing-notes.md, no licence_status, no ATTRIBUTION.txt existed before this pass — LU had never been through this project's proper rebuild.

Verdict

Build it. No licensing obstacle, no coverage gap.

Build — 2026-09-02 (local only, not published)

Built via datasets/lu/scripts/build_postal_codes_product.py.

Frictionless validation: valid: True. Zero literal "nan" strings, zero float-formatted accuracy values, zero duplicates on the real primary key (postal_code, place_name, admin_name2).

Status: built and validated locally only. Nothing published to R2, nothing committed to git. datasets/lu/README.md and datapackage.yml updated in place (tier B, premium_source_key: lu/lu.zip, real sources/licenses/licence_status, replacing the stale "100-row sample" placeholder text).