[MaxGhenis]

This PR adds three Belgian sources for Microcosm-BE's transport port (PolicyEngine/microcosm#1369, decision d048).

Inputs they feed (method card BE-MC3, pending Max's ruling d1902):

• The USD→EUR conversion of donor money.
• The quantile map of legal pensions onto Belgian amount bands.
• Pension caseload targets by region and commune.

Each package is a source-only facts package. Every published cell becomes one fact, and nothing is derived, rescaled or aged.

Packages

Package	Facts	Source
eurostat-ert-bil-eur-a-usd (new)	55	Eurostat ert_bil_eur_a: annual average USD per euro/ECU, 1971–2025. 2024 is 1.0824.
sfpd-pension-amount-bands-2025 (new)	10,239	FPD Jaarlijkse statistiek 2025, sections 1.2, 6.2 and 9: January 2025 recipients and monthly expenditure by gross amount band, sex, career, category, age and IGO eligibility.
sfpd-legal-pension-caseload-2025 (extended)	20,954	PensionStat 2025 workbook (CC0). Pensioners, commune caseloads, IGO counts, averages and medians, printed annual expenditure, and minimum pensions. These replace the package's four curated webpage headlines, which are not cells in the workbook.

GPT-6.1 Sol lanes built each package. Each lane passed validate-package, build-suite and focused tests, plus independent ledger-source-fidelity and ledger-boundary reviews. The pension-band lane also re-read the PDF with Poppler and matched all 10,239 values and locators.

Changes made at integration

PensionStat communes. The lane matched 511 of 565 commune names and kept 54 as publisher-local scopes. Those 54 were mostly Dutch names of Walloon and Brussels communes: Aarlen, Luik, Elsene and so on.

Matching against Statbel's official Dutch and French names (now on main through #383) and the publisher's own French edition assigns all 565. The generator fails unless the result is a bijection onto Statbel's 565 NIS 2025 communes. The tiers:

• Statbel Dutch name: 551 exact and 9 orthographic (case, diacritics, apostrophe style, œ).
• French edition: 4 communes. Each commune of the Dutch edition is paired with the French-edition row that has the identical 2019–2025 count series; all 565 series are distinct. The French label then matches Statbel's French name. This resolves Belle (Belœil), Gembloers (Gembloux), Heilissem (Hélécine) and 's-Gravenbrakel (Braine-le-Comte).
  • The French workbook is registered as label evidence (label_evidence_fr) and yields no facts.
• One declared section label: PensionStat calls Saint-Nicolas (Liège province, 62093) by its section Montegnée in both editions, because its French edition labels Sint-Niklaas "Saint-Nicolas".
  • The generator declares that collision and fails if it stops holding.
  • 62093 is also the only commune left without a row once the other 564 are assigned.

commune_mapping.json records each commune's tier, its French-edition label and cell, Statbel's names and the source cell.

Geography names. Country, NUTS1 and commune names follow the existing Statbel packages ("Belgium", "Flemish Region", "Arlon"). The publisher's labels stay in each package's own dimensions. The lanes had used "België" and the Dutch region names, which would have added four cross-package name warnings.

Period coverage of counts (from the first review). Both publishers date some counts to a stock on 1 January, not to the month:

• The FPD tables' section headers read "Toestand op 1 januari 2025" (PDF pages 25, 225 and 239).
• PensionStat's IGO page says "De aantallen hebben betrekking op 1 januari 2025".

So the 5,572 FPD recipient counts and 1,715 IGO counts now have coverage from 1 January 2025 to 1 January 2025. Monthly expenditure and the IGO mean and median stay at 1–31 January.

The other PensionStat sheets keep January coverage, because their pages count everyone who received a pension in January 2025. Those sheets are pensioners, communes and minimum pensions.

Coverage is declared per record set, so:

• each FPD band group is now one record set per metric;
• the IGO sheet is two record sets, counts and amounts.

The split sets share their table's spec ID and source-record-ID prefix. All 31,193 fact keys, values, periods and source-record IDs are therefore unchanged, and so is every bundle pin.

What moves, for the 15,384 facts in the split tables:

• layout.record_set_id and the layout spec hash;
• the measure ordinal, for the 8,097 facts whose measure moved into a new set (FPD expenditure, IGO mean and median);
• the coverage notes;
• the coverage end date and source label, for the 7,287 counts only.

Rebase onto main. The lanes branched from 8f4cfd5, before #383 and #384. All bundle pins were re-derived on 200bf6d. Notably, chronicle#381 already added calendar years 1991–1994, so the exchange-rate series adds 18 period keys, not 22.

Merged-bundle pins

The merged bundle changes as follows:

• Facts: 1,586,258 → 1,617,502 (+55 +10,239 +20,954 −4 retired).
• Sources: 57 → 58.
• Packages: 413 → 415.
• Tables: 355 → 357. The PensionStat workbook table replaces the curated one.
• Geographies: 102,613 → 102,615 (Eurostat's USD scope and PensionStat's residents abroad, BUI). The 565 commune keys already exist through Add Statbel commune population facts for NIS 2025 #383.
• Periods: 535 → 553.
• Warnings: 123 (+1). Eurostat's exchange-rate dataset labels freq=A "Annual", as its JSON-stat file prints it, where the seven older Eurostat packages carry the bare code A. That adds one conflicting_value_label_across_packages warning, pinned in 49809b6. The first CI bundle run at 0a6f053 caught it; it was the only failing assertion.

Per-key pins are updated for country:BE, nuts1:BE1..3, commune:11002, entities, sources and periods. New source-only coverage tests in tests/test_chronicle_bundle.py record the measurement behind each pin.

Invariants (tested)

• One fact per published cell. Every numeric 2025 cell in the selected sheets, sections and series becomes exactly one fact. Values are verbatim, with no scaling, ratios or totals.
• Communes. PensionStat's commune rows map onto NIS 2025 as a bijection (565 → 565). Geography names equal the Statbel canonical names.
• Determinism. Each generator reproduces its checked YAML byte for byte, with no YAML anchors or merge keys.
• Coverage follows the publisher's reference date. Counts dated "Toestand op 1 januari 2025" or "betrekking op 1 januari 2025" cover that day. Monthly amounts cover January. Split record sets keep one table identity.
• Raw captures. Each capture's SHA-256 and size match its manifest. R2 keys are content-addressed under ledger-raw.

Checks

Check	Result
test_chronicle_pensionstat.py	12 passed locally at f65e629; passed at 0a6f053 in CI (Chronicle surface)
test_chronicle_sfpd_pension_amount_bands.py	20 passed locally at f65e629; passed at 0a6f053 in CI (Chronicle surface)
test_source_eurostat_exchange_rates.py	13 passed
test_belgium_targets.py	26 passed
test_chronicle_bundle.py -k source_coverage	2 passed
Fact identities before and after 0a6f053 (all 31,193 facts)	keys, values, periods and source-record IDs identical; coverage dates changed for exactly the 7,287 counts
validate-package / build-suite, each package	valid, 0 errors, 0 warnings
Per-source bundle checks (dimension labels, geography names)	0 issues
ruff check chronicle db scripts tests (0.14.10)	pass
CI Chronicle surface at 0a6f053 (every test file except the bundle, UK national and OpenTimestamps files)	pass

The merged-bundle test was not run locally (about 15 GB peak, about 2.5 hours). CI's bundle job is its measurement. The test stops at its first failed assertion, so the 0a6f053 run never reached its later pins. Those were checked by replaying the test against that run's uploaded bundle reports: the replay fails at 0a6f053 exactly where CI did, catches deliberately broken pins, and passes at 49809b6.

All five raw files are in R2 ledger-raw under their content-addressed keys, each read back and matched by SHA-256.

Review. The independent GPT-6.1 Sol review of f65e629 confirmed every bundle pin, the commune bijection and all sampled cells. It requested the two coverage fixes above.

Its re-review of 0a6f053 approved with no findings. An independent Opus 5.5 review of the 49809b6 pin commit also approved; it checked the warning against all eight Eurostat raw files and the artifact. That commit also sorts an import, a cosmetic change: the repo's ruff config does not select the import-order rule, despite its commit message. It checked each sheet's coverage against the publisher's definition pages, and independently rebuilt both revisions' facts: keys, values, periods, measures, source-record IDs and lineage are identical for all 31,193 facts.

🤖 Generated with Claude Code

[MaxGhenis]

Review of f65e629 (GPT-6.1 Sol, REQUEST_CHANGES with two P2 findings), addressed in 0a6f053:

1. FPD recipient counts dated to 1 January. The 5,572 FPD recipient counts now cover 2025-01-01 to 2025-01-01, matching "Toestand op 1 januari 2025" on PDF pages 25, 225 and 239. Monthly expenditure keeps January.
2. IGO counts dated to 1 January. The 1,715 IGO counts likewise cover 2025-01-01 to 2025-01-01, matching "De aantallen hebben betrekking op 1 januari 2025". IGO means and medians keep January.

Record sets are split per coverage but share each table's spec ID and source-record-ID prefix. A before/after dump of all 31,193 facts shows identical aggregate, semantic and legacy keys, values, periods and source-record IDs, so the bundle pins are unchanged. I checked the other PensionStat sheets against their definition pages: pensioners, communes and minimum pensions count recipients "in januari 2025", so they keep January coverage.

The tests now restate the publisher dates independently and check the PDF headers. The re-review is running, and the focused suites are queued behind tonight's host reservation.

[MaxGhenis]

Merge audit for head 0a6f0538271b87c109136b64c444ff35a9e66f79.

• Independent review: GPT-6.1 Sol re-review of 0a6f053 approved with no findings. It rebuilt both revisions' facts, and all 31,193 are identical in key, value, period, measure, source-record ID and lineage. The earlier review of f65e629 asked for the two coverage fixes this head contains.
• Checks: Chronicle surface passed at this head. Chronicle bundle is running. A watcher merges with squash, --match-head-commit 0a6f0538, only when gh pr checks exits 0 and the PR is MERGEABLE, not a draft and has no CHANGES_REQUESTED. It aborts if the head moves.
• If the only block is the required approving review, the watcher admin-merges at this head under the standing rule for the owner's own PRs. Failing or pending checks and conflicts are never bypassed.
• Shared code: only SOURCE_PACKAGE_ALIASES registration lines and the bundle pins for these packages.

🤖 Generated with Claude Code

[MaxGhenis]

Audit, 49809b6: the CI bundle run at 0a6f053 failed on one pin, warning_count 123 against the pinned 122.

The extra warning is conflicting_value_label_across_packages for freq=A. Eurostat's exchange-rate dataset labels it "Annual", as its JSON-stat file prints it. The seven older Eurostat packages carry the bare code. 49809b6 pins the warning and leaves both labels as published.

The test stops at its first failed assertion, so CI never reached the later pins. I checked them by replaying the full test against that run's uploaded bundle reports. The replay fails at 0a6f053 where CI did and passes at 49809b6.

Independent review of 49809b6 (Opus 5.5): APPROVE. Its one request, the stale warning count in the PR body, is fixed. Earlier reviews: GPT-6.1 Sol approved 0a6f053 with no findings; that review is quoted in the PR body. This PR merges (squash) only when CI is green at 49809b6.

🤖 Generated with Claude Code

[MaxGhenis]

Merged origin/main (#386, ACS county packages) in a8f68f0.

• The conflict: two source-package pins in tests/test_chronicle_bundle.py. The resolution keeps 415 built packages (this PR) and takes main's 24 skipped packages; Add ACS 2020–2024 county income source packages #386's seven county packages have no 2023 artifact. The diff from the reviewed head 49809b6 is exactly Add ACS 2020–2024 county income source packages #386's diff from 200bf6d.
• CI at 49809b6: fully green, including the 2h47m bundle test.
• Merge review (GPT-6.1 Sol, Subfleet run 20261024-020920-chronicle385-merge-review): APPROVE at a8f68f0. ledger-source-fidelity: PASS; ledger-boundary: PASS. It found that the merge adds nothing beyond main, and that Add ACS 2020–2024 county income source packages #386's skipped packages and its suite.py change cannot affect any of this PR's pins or packages.

This will squash-merge at a8f68f0 once CI passes on this head.