Context
PolicyEngine/policyengine-us-data#1151 exposed a class of failure where support/clone rows can dominate final Enhanced CPS weights even when headline household totals look plausible. Microplex uses PUF and other donor sources differently, but the same production risk exists: a candidate can pass target-fit checks while final calibrated mass is concentrated in a support source or clone-like donor subset.
Microplex should keep source/provenance diagnostics out of the exported PolicyEngine H5, but we still need a separate artifact and release gate for this.
Proposed Microplex behavior
- During canonical mp build/scoring, write a sidecar source-weight diagnostics artifact outside the H5.
- Summarize final calibrated mass by construction source or source class, at least:
- household count and weight share
- person count and weight share
- tax-unit count and weight share where available
- positive-weight row counts
- top source/source-class concentration
- Include PUF-specific diagnostics where supported by construction metadata:
- original CPS/ASEC-like rows vs PUF donor/support rows
- any explicit support-clone or donor-replay subset
- top-tail/fixed-spine rows as a distinct fixed source, not mixed with ordinary PUF support
- Add an mp artifact gate that fails when source/support rows dominate beyond configured thresholds unless explicitly marked as an experimental build.
- Keep all source/provenance fields out of the exported PolicyEngine model variables; diagnostics live in sidecars and manifests only.
Acceptance criteria
- A canonical build writes a stable JSON sidecar, e.g.
source_weight_diagnostics.json, referenced from manifest.json.
mp300k_artifact_gates consumes the sidecar and reports a source_weight_diagnostics gate.
- The gate is required for replacement candidates and unmeasured/failing if the sidecar is missing.
- Tests cover a passing balanced-source payload and a failing clone/support-dominance payload.
- Export guards continue to reject source diagnostic variables in the H5.
Related us-data precedent: PolicyEngine/policyengine-us-data#1151
Context
PolicyEngine/policyengine-us-data#1151 exposed a class of failure where support/clone rows can dominate final Enhanced CPS weights even when headline household totals look plausible. Microplex uses PUF and other donor sources differently, but the same production risk exists: a candidate can pass target-fit checks while final calibrated mass is concentrated in a support source or clone-like donor subset.
Microplex should keep source/provenance diagnostics out of the exported PolicyEngine H5, but we still need a separate artifact and release gate for this.
Proposed Microplex behavior
Acceptance criteria
source_weight_diagnostics.json, referenced frommanifest.json.mp300k_artifact_gatesconsumes the sidecar and reports asource_weight_diagnosticsgate.Related us-data precedent: PolicyEngine/policyengine-us-data#1151