EPIC: populace-be — Belgium as the first Axiom-engine country

[MaxGhenis]

Goal

Ship a working Belgian tax-benefit model — "policyengine-be" as a product outcome — composed as Axiom rules × populace data, with no policyengine-be country package. Belgium is the first country built on the Axiom engine through the Frame RulesEngine protocol, and the first country expressed entirely as declarative specs. The build recalibrates the proven US release recipe to Belgian sources and targets.

Why Belgium first

• Rule encodings for Belgium are starting on the Axiom side; the populace half can be scoped to the encoded slice and grow with coverage.
• Belgium publishes unusually good open calibration aggregates, including fiscal income statistics at the municipality level (Statbel), social-security aggregates (ONSS/RSZ), and caseload series — a strong target surface without licence friction.
• The base microdata (BE-SILC) is licence-restricted the same way the UK FRS is, so the private-artifact + public-diagnostics machinery transfers wholesale.
• EUROMOD-BE (open access) and Federal Planning Bureau published costings provide external validation oracles, replacing the self-parity gate that first-party countries used.

Binding architecture constraints

1. Spec-first (greenfield consumer of Move country build choices into declarative specs #159 and Consume Ledger target profiles instead of carrying target values in Populace #160). No be_runtime/ Python. Country content = spec files + Ledger source packages/target profiles. Anything code-shaped becomes a general, parameterized, tested operator in the core packages — never country Python.
2. Rules only via the adapter. All engine access goes through packages/populace-frame/src/populace/frame/rules.py (RulesEngine protocol); the Axiom adapter is the only module that imports the engine. Reference implementation: packages/populace-frame/src/populace/frame/adapters/policyengine_us.py.
3. Licence posture. BE-SILC microdata never leaves a private HF repo; destination guards land before any upload code (mirror policyengine-uk-data utils/hf_destinations.py + AST-scan test). Public artifacts are aggregates only: manifests, calibration diagnostics, source coverage, reform validation.
4. Naming. data_build_id convention (builder + build sha + date); no ordinal versioning.

Phases

• Phase 0 — scope. Confirm the Axiom BE coverage slice (assume PIT + social contributions first); confirm SILC access path and institutional holder; inventory targets.
• Phase 1 — foundations. ledger-be source packages + target profiles; spec schema v0 with BE as first consumer; Axiom RulesEngine adapter (both sides).
• Phase 2 — build. SILC → Frame source stage under the licensed environment; clone-and-assign geography spine (commune NIS codes, vintage-aware); calibration to the Belgian surface with the US recipe.
• Phase 3 — prove and ship. EUROMOD-BE/BFP validation; release channel; dashboard country toggle.

Children

Foundations (phase 1)

• Implement an Axiom-backed RulesEngine adapter #260 — Axiom-backed RulesEngine adapter (consume side); provide side: Expose a vectorized engine surface for the populace RulesEngine adapter (Belgium pilot) TheAxiomFoundation/axiom-rules-engine#62
• Define the declarative country spec schema with Belgium as its first consumer #261 — declarative country spec schema, Belgium as first consumer (realizes Move country build choices into declarative specs #159 + Consume Ledger target profiles instead of carrying target values in Populace #160)
• Author ledger-be source packages for open Belgian aggregates with geography vintages ledger#69 — ledger-be source packages (open Belgian aggregates, NIS geography vintages)
• Define Belgian target profiles for populace-be ledger#70 — Belgian target profiles
• Facts-only store: type source-published projections as facts, enforce period contracts at consumption ledger#71 — facts vs projections schema split (aging; also resolves the Period-aware transformation needed for SOI EITC-by-AGI calibration targets #116/Cross-check: Populace US 2024 net income & tax ~37–43% below Enhanced CPS (benefits match) #212 residual)

Build (phase 2)

• Add the BE-SILC source stage with licence guardrails and a private artifact repo #262 — BE-SILC source stage + licence guardrails + private artifact repo
• Build the Belgian calibration surface by recalibrating the US release recipe #263 — Belgian calibration surface (US recipe recalibrated; commune spine, vintage-aware)

Prove and ship (phase 3)

• Validate populace-be against EUROMOD-BE and Federal Planning Bureau scores #264 — validation oracles: EUROMOD-BE baseline + Federal Planning Bureau reform scores
• Stand up the populace-be release channel #265 — release channel (contract artifacts, private HF, latest.json, staging)
• Add Belgium as a third country toggle calibration-diagnostics#62 — Belgium country toggle

Dependency notes: #263 depends on #260, #261, #262, ledger#70. #264 depends on #260 (reform materialization decision) and #263. #265 depends on #263. calibration-diagnostics#62 depends on #265. ledger#70 depends on ledger#69 and ledger#71 (projection-basis targets).

Out of scope (follow-ups)

• policyengine.py registration for BE (requires Axiom-engine support in the .py stack; separate epic once the artifact exists).
• Firm/enterprise population for BE (the UK firm module pattern generalizes later).
• A public synthetic BE artifact (requires the disclosure criteria from DESIGN.md).

Acceptance criteria

• A populace_be release (private artifact) passes the full release contract: calibration, target_profile_coverage, base_population_scale, demographics-vs-Statbel gates, with reform validation against named external oracles on the encoded slice.
• Public diagnostics render in calibration-diagnostics under a be country toggle.
• The BE country folder contains zero Python.
• The Axiom adapter passes the same behavioral contract tests the policyengine-us adapter passes (where semantics overlap).

[MaxGhenis]

Phase-1 foundations status (2026-07-02):

Landed

• Dense Python surface for the populace RulesEngine adapter TheAxiomFoundation/axiom-rules-engine#63 (merged) — the vectorized engine surface for Implement an Axiom-backed RulesEngine adapter #260's provide side: fixed the dense PyO3 extension build (missing DenseColumn::Float arm; CI now cargo-checks the crate), exposed execute_f64, and added derived_metadata (entity/dtype/period/unit/source per derived rule). Verified on rulespec-be: article 130 reproduces hand-computed 2025 liabilities; the full federal PIT pipeline (49 outputs) materializes 200k persons in ~1s. Remaining engine-side follow-ups (typed input specs, Count dtype alias, coverage inventory) are itemized on Expose a vectorized engine surface for the populace RulesEngine adapter (Belgium pilot) TheAxiomFoundation/axiom-rules-engine#62.

In review (CI green)

• Axiom-backed RulesEngine adapter: Belgium as the first Axiom-engine country #268 → Implement an Axiom-backed RulesEngine adapter #260: populace.frame.adapters.axiom.AxiomEngine, the first non-PolicyEngine RulesEngine adapter, with BE_SCHEMA (person+household), the adapter-owned entity-table H5 format (AxiomEntityTableDataset, US/UK single-year layout), a strict formula-owned export gate, and the behavioral contract suite parameterized over adapters (policyengine-us + Axiom pass the same file). Decisions the issue asked for are in the PR body — notably reform materialization: no protocol extension; a counterfactual is a second adapter over a reform-compiled module, so Validate populace-be against EUROMOD-BE and Federal Planning Bureau scores #264 sequences behind reform modules compiled upstream.
• Declarative country spec schema with Belgium as its first consumer #269 → Define the declarative country spec schema with Belgium as its first consumer #261: the declarative country spec (populace.build.country_spec) with Belgium as first consumer — be/ country package (zero Python, auto-enforced): BE-SILC source stage with the income-reference-year lag declared as data, vintage-aware commune NIS spine (mismatch = compile error), six target-reference families by reference only (values-in-populace structurally refused), gate selection with criticality tiers (incumbent-comparison gates deliberately unselected — external oracles replace self-parity), and the private-repo release contract. Golden-file test pins every spec byte.

Next up (unblocked)

• Add the BE-SILC source stage with licence guardrails and a private artifact repo #262's blocking first PR: the HF destination allowlist + AST-scan guard (the UK licensed-data pattern), before any BE upload code exists.
• ledger#69/Investigate ACA take-up and plan-choice inputs driving high PTC estimates #70/Guard US fiscal refresh against degenerate ACA inputs #71 (BE source packages, target profiles, facts-vs-projections) — the be/target_references.json selectors name the intended package ids so profile binding needs no populace-side changes.

Blocked externally: BE-SILC access path / institutional holder (phase 0 scoping from IMA), which gates the #262 source stage itself and everything behind it.