Author ledger-be source packages for open Belgian aggregates with geography vintages

[MaxGhenis]

Part of the populace-be epic PolicyEngine/populace#259. Sibling: the BE target-profiles issue.

What

Author ledger-be source packages for the open Belgian aggregate series that will drive populace-be calibration and validation. Pure public statistics — no licence friction. Each is a source_package.yaml + declarative selectors under packages/ following the agent-source-package harness and governance rules.

Initial package list (verify exact series/table names against the publishers during authoring — the validate-package harness gates):

• Statbel — fiscal statistics of income by municipality (statistique fiscale des revenus / fiscale statistiek van de inkomsten): total net taxable income, filer counts, income distribution per commune. The flagship subnational series.
• Statbel — population structure: age × sex × commune/region, reference-date population.
• ONSS/RSZ: employment and social-contribution aggregates by sector/region.
• ONEM/RVA: unemployment benefit caseloads.
• SFPD: pension caseloads and expenditure.
• Regional child benefit (post-2019 regionalization): caseload/expenditure per region (Groeipakket and counterparts).
• NBB national accounts: household-sector income levels (validation-tier anchor).
• Federal Planning Bureau: economic outlook series (the OBR-analog for projections) + published reform costings referenced by the validation oracle.
• EUROMOD-BE country statistics: published baseline aggregates used as comparators.

Geography requirement (blocking for subnational packages)

Commune NIS codes change across years (municipal mergers). Every subnational fact carries geography vintage metadata, same mechanism as the congressional-district vintage work (PolicyEngine/populace#205): facts declare their code-set vintage; consumers get a compile error on vintage mismatch rather than a silent partial join. A NIS-code vintage crosswalk (old→merged communes) is part of this issue.

Acceptance

• Packages pass validate-package; facts load with full lineage to source cells.
• Subnational facts all carry geography vintage; the crosswalk exists for at least the most recent merger wave.
• The sibling target-profiles issue can bind these facts without any populace-side value copying.

[MaxGhenis]

Execution brief (2026-07-02) — everything needed to build this now

Phase-1 populace-be infrastructure landed since this issue was filed: the Axiom adapter (populace#268), the declarative be/ country package (populace#269), and live EUROMOD-BE per-case oracles (TheAxiomFoundation/axiom-oracles#73, merged). Ledger is now the critical-path gap for a working Belgium. This comment is the self-contained brief for the agent building the packages.

The binding contract: populace selects by these source names

populace's packages/populace-build/src/populace/build/be/target_references.json (PR populace#269) binds targets by ledger_selector — these source_names are the interface; package ids/selectors must resolve them (or the same PR that changes a name must update the populace spec):

source_name	geography_level	populace measure it calibrates	publisher
statbel_population_structure	nuts1	people (age band × sex × region)	Statbel
statbel_fiscal_income	commune (nis_vintage 2025)	belgium_pit_taxable_income	Statbel
spf_finances_pit	country	belgium_pit_federal_and_local_tax_before_withholding (assessed PIT: federal after art. 134 + arts. 465–468 local)	SPF Finances / Statbel
onss_contributions	country	belgium_worker_article_17_uncapped_component_contribution	ONSS/RSZ
onem_rva_unemployment	country	receives_unemployment_benefit (caseload)	ONEM/RVA
nbb_national_accounts	country	household_disposable_income (validation-tier band, never hard-calibrated)	NBB

Package list beyond those six (from the issue body): SFPD pension caseloads/expenditure, regional child benefit (Groeipakket + FWB/German-speaking counterparts, per region), BFP economic outlook (the OBR-analog for projections), and published EUROMOD-BE baseline statistics (the #264 comparator package).

Publisher entry points (verify exact tables/series during authoring — the validate-package harness gates)

• Statbel population structure: statbel.fgov.be → Population → Structure of the population (reference-date population by age/sex/municipality; NUTS1 rollup).
• Statbel fiscal statistics of income: statbel.fgov.be → Households → Taxable income (per-municipality net taxable income, filer counts, distribution).
• ONSS/RSZ: onss.be statistics / socialsecurity.be open data (employment + contribution aggregates).
• ONEM/RVA: rva.be → statistics (unemployment benefit recipients).
• SFPD: pensionstat.be (caseloads/expenditure).
• NBB: stat.nbb.be (household-sector accounts).
• BFP: plan.be (economic outlook; reform costings for the oracle list).
• EUROMOD-BE: JRC country reports + released baseline statistics.

Geography vintage (blocking for subnational packages)

Every commune-grain fact carries NIS code-set vintage metadata; the populace spine declares vintage 2025 with vintage_policy: error (populace#269 geography_spine.json), so a mismatched vintage must be a compile error, never a silent partial join (the populace#205 lesson). Author the old→merged crosswalk for at least the 2025-01-01 merger wave (Statbel publishes the NIS code list + fusion mappings); the crosswalk fits #71's derived-transform conventions.

Basis periods (coordinate with #71, don't block on it)

Author facts strictly as published, with explicit periods (fiscal income for income year N is published per assessment year N+1; SILC-based comparisons lag a year). Where populace needs a different basis (e.g. PIT total aged to the simulation year), that's a projection under #71 — never a silently re-dated fact. If #71's schema isn't merged when you get there, leave projection-basis targets to a follow-up rather than smuggling aged values into facts.

Acceptance (agent-checkable)

1. Every package passes validate-package; facts load with full lineage to source cells.
2. All subnational facts carry nis_vintage; the 2025 crosswalk exists and round-trips a merged commune.
3. A resolution test proves each of the six ledger_selectors above resolves to exactly one fact stream (this is what populace#269's spec binds to).
4. The EUROMOD-BE comparator package exists with source URLs per row (feeds populace#264's reform_validation.json).
5. No observed values copied anywhere outside Ledger (populace consumes by reference only — structurally enforced on the populace side).