Great Britain (GB): migration notes

Internal record of the P5 migration from the old scripts to the producer: the discrepancies found, their causes, and who decided what. This is a dated record, so the numbers are as of the migration and not kept current. It is not shipped in the zip.

• Issue: pc-28e.7.23 (closed 2026-10-03)
• Commit: 20284af (worktree commit 4c32157), landed in batch 5
• Parity verdict: unexplained (the gate blocks on any unclassified cell); accepted as upstream drift plus two user decisions
• Producer: custom, refresh disabled

Sources

• The old scripts read hand-fetched files from sources/; no download URL was recorded. The agent found the OS Downloads API URL and the ONS item ids through the ArcGIS search API.
• The ONS lookups are fixed vintages, so their URLs don't move. A future Code-Point release on newer ward codes makes the producer stop with "ward code … isn't in the ONS lookups" until someone adds that vintage to WARD_URLS.

Parity against the live zip (2026-10-03)

Rows: 1,747,841 live and 1,749,109 fresh. 2,237 only in live, 3,505 only in fresh. 157,082 rows changed, 332,536 cells, all unclassified. Line endings change from CRLF to LF.

Changed cells by column: admin_code3 120,103; admin_name3 59,577; latitude 37,802; longitude 38,087; easting 36,833; northing 36,726; place_name 2,874; accuracy 362; admin_code2 / admin_name2 68 each; admin_county_code / admin_county_name 16 each; admin_code1 / admin_name1 2 each.

How the agent showed the logic reproduces the old build:

• On every row whose easting/northing is unchanged, latitude/longitude match exactly (the pyproj BNG→WGS84 conversion matches).
• On every row whose GSS codes are unchanged, the district, ward and county names match (lookups and ward load order match).

1,659 + 562 + 653 = 2,874 place_name cells.

Decisions

Transforms carried over

Folded from parse_codepoint.py → enrich_uk.py → fill_place_name.py → data_integrity_checker.py (restored 2026-08-25):

• Rows from every Data/CSV/*.csv in the Code-Point zip, in sorted file order; cells stripped.
• admin_code1 = CY, mapped to England / Scotland / Wales; admin_code2 = DC; admin_code3 = WC; admin_county_code = CC; accuracy = PQ; easting/northing verbatim.
• BNG (EPSG:27700) → WGS84 with pyproj, 6 decimals. Positional-quality 90 rows (0,0) get blank coordinates.
• Names joined on GSS code from the ONS lookups; the ONS UTF-8 BOM is stripped.
• place_name = nearest GeoNames populated place from the 6-decimal coordinate; rows without a coordinate have none. alternative_city_name is always empty.
• Old checker gates are ProducerErrors: postcode pattern, unique postcode, known country code, GB bounding box. New: all three countries present, every district/ward/county code resolves.

Open questions

• None recorded from the agent beyond the two answered above.
• ATTRIBUTION.txt still has counts ("1,747,841 unit postcodes", "865 of 1,747,841", "31.9% of rows"), and .licensing-notes.md has historical counts. Left alone; they belong to the pc-hfz sweep.

Provenance

Reconstructed 2026-10-05 from the agent transcript d33e1b7e-5027-498a-a97f-ba914ee2524e/subagents/agent-ada3531935232a997.jsonl, the orchestrator session d33e1b7e-… (AskUserQuestion answers and close), the commit message, the bd close reason and the batch 5 sign-off in docs/parity/p5.md.