Add Belgium pilot: Axiom rules engine over populace-be entity tables

[MaxGhenis]

Fixes #447

What this is

A pilot Belgium channel for policyengine.py — the first country that runs on the Axiom rules engine instead of a policyengine-core country package:

• src/policyengine/tax_benefit_models/be/ — AxiomBelgiumPilot, a TaxBenefitModelVersion subclass (deliberately outside the certified-release machinery: no manifest, no region registry) whose run() builds a populace Frame, executes the rulespec-be composed worker pipeline through the populace AxiomEngine adapter, and returns weighted person/household MicroDataFrames as the output dataset.
• PopulaceBelgiumDataset / BEYearData — two-entity (person, household) HDF5 dataset in the populace-be layout, calibrated household weights.
• examples/belgium_axiom_pilot.py — the end-to-end population run.
• tests/test_be_axiom_pilot.py — statutory assertions on a four-person fixture around the work-bonus phase-out (contribution wiped at 20k, partial at 30k, statutory 13.07% by 60k); skips cleanly unless populace-frame, axiom-rules-engine, and a rulespec-be checkout are present (source-only, so CI skips).

Validation status: machine-checked conformance vs EUROMOD

The encoding this channel executes is conformant against EUROMOD BE_2025 per the axiom-oracles scoreboard: 23/23 in-scope EUROMOD policies covered, 0 unexplained residuals, 0 open Axiom-attributed mismatches. Per-case, the two engines agree to the cent on the worker suites — including be-worker-disposable-income-list (gross − tax − SSC end to end, 0 mismatches) — and where they deliberately differ (e.g. the couple pipeline scopes out the worker-stage deductions), each residual carries a committed disposition with a statutory citation. Compositions behind every covered suite are now machine-recorded and reproducible (TheAxiomFoundation/axiom-oracles#235). Nothing in the tables below is unexplained.

Population results on the calibrated pilot dataset (2026-07-08)

populace-us support records reweighted to Belgian Ledger targets (Statbel demography, SPF fiscal income, ONSS contributions, ONEM recipients — 21/21 within 1.8%), 166,302 persons:

aggregate	this channel (Axiom)	EUROMOD, identical rows
employee SSC (gross-of-bonus concept)	EUR 20.874B	EUR 20.874B — agreement to EUR 643
employee SSC after work bonus	EUR 16.16B	EUR 15.02B
worker PIT before withholding (individual assessment)	EUR 23.28B	EUR 19.53B

Reading the PIT difference — every euro is attributed, and the buckets point in different directions:

1. EUROMOD-attributed (~EUR 2B, documented and raised upstream). EUROMOD refunds the fiscal work-bonus credit without capping it at contributions due; CIR 92 article 289ter/1 credits the bonus "reellement accorde", which ONSS caps at contributions — the encoding follows the statute and the divergence is logged oracle-attributed in the conformance ledger. The work-bonus phase-out/January-timing residual is filed on the JRC's own engine repo (Belgium BE_2025 employee work-bonus output appears to apply February 2025 table to January ec-jrc/JRC-EUROMOD-software-source-code#5). EUROMOD also does not model the German-speaking community regime at all (Axiom does, grounded to the DG decrees).
2. Axiom-side composition, tracked (~EUR 2B when couples/kids are linked). Dependant allowances and the marital quotient are encoded and their residuals dispositioned, but the couple pipeline does not yet chain the worker stages (SSC/forfait/work-bonus) per spouse — Chain worker SSC/forfait/work-bonus stages into the couple PIT pipeline (retire the marital-quotient dispositions) TheAxiomFoundation/rulespec-be#106 tracks that composition (compositions themselves are now mechanically recorded and resolvable per conformance: record per-suite runnable program compositions (#185) TheAxiomFoundation/axiom-oracles#235). With units linked, EUROMOD computes EUR 17.15B; closing that distance is on this side, not theirs.
3. Demo-population labeling artifact. The pilot labels every worker full-time/full-year, which inflates low-wage bonus divergence on both engines' FTE machinery; this washes out with real data (SILC carries actual months and hours).

On the SPF Finances comparison in earlier revisions: worker PIT sits at 0.37 of SPF's EUR 62.84B all-PIT total — and moved further from it as the encoding improved (the newly encoded refundable credit reduces worker-slice PIT). That distance is scope by construction, not error: the pipeline models wage earners only (EUR 159.7B of remuneration against EUR 274.7B of fiscal income; no pensions, self-employed, or investment income in the base).

What this is not

Not a certified model and not representative Belgian microdata — the support records are American, worker-slice coverage, individual assessment. Labeled as such throughout. It demonstrates the channel (Simulation(dataset=..., tax_benefit_model_version=AxiomBelgiumPilot(...)).run()) so the certified pieces (SILC support strata, engine-computed targets, full composed coverage) can land behind a stable surface.

Dependency note

populace-frame and axiom-rules-engine are not on PyPI; nothing is added to pyproject.toml. The module docstring and test skip conditions document source installation. Publishing both is the follow-up that makes the example runnable for users (axiom-rules-engine wheel publishing is blocked on a maturin build regression on its main, filed as TheAxiomFoundation/axiom-rules-engine#92).

🤖 Generated with Claude Code