Fix mp-300k target support on HoH AGI, SSI, Medicare, and ACA benchmark families

[MaxGhenis]

Parent: #11

Current-main validation after the table-construction fixes shows the first small replacement candidate is no longer primarily failing because of tax-unit/SPM/family fragmentation. Entity structure is now close to eCPS on the measured matched-N diagnostics, but the sound eCPS replacement comparison still fails by a wide margin.

Evidence artifact:

/Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/sound_ecps_replacement_comparison/sound_ecps_replacement_comparison.json

Manual per-target drilldown:

/Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/sound_ecps_replacement_comparison/target_diagnostics_refit_ecps_to_mp.json

Current loss result:

• Microplex refit PE-native broad loss: 3.3933854903630642
• eCPS refit PE-native broad loss: 0.16459032249699052
• Microplex train/holdout: 2.88954251907915 / 0.5038429754602355
• eCPS train/holdout: 0.1372832454307687 / 0.027307076904983546
• target wins: eCPS 2293, Microplex 496, ties 29

Entity structure after #79/#80/#83 is not the main gap anymore:

• Microplex matched tax units/HH: 1.3079; eCPS: 1.3377
• Microplex matched SPM units/HH: 1.0000; eCPS: 1.0441
• Microplex matched families/HH: 1.0928; eCPS: 1.1188
• Microplex family singleton share: 0.3787; eCPS: 0.4148
• cross-household family units: 0 for both

Largest target blockers after refit include:

• Head-of-Household IRS AGI/count bins, e.g. AGI 20k-25k, 25k-30k, 500k-1m, and 1m-inf taxable HoH cells.
• Medicare Part B premium targets by young ages, where eCPS is often zero and MP has large positive estimates.
• SSI recipients and SSI dollar totals, especially age 65+ and CBO SSI.
• State ACA spending and SNAP household families.

The current target diagnostics show examples like:

• nation/irs/count/count/AGI in 20k-25k/taxable/Head of Household: target 53,987, eCPS 22,976, MP 1,131,244
• nation/irs/count/count/AGI in 25k-30k/taxable/Head of Household: target 131,400, eCPS 200,375, MP 2,633,483
• nation/irs/count/count/AGI in 500k-1m/taxable/Head of Household: target 43,007, eCPS 42,528, MP 858,975
• nation/census/medicare_part_b_premiums/age_10_to_19: target 11.2M, eCPS 0, MP 222.9M
• nation/ssa/ssi_recipients/65_plus: target 2.38M, eCPS 3.81M, MP 32.13M
• nation/cbo/ssi: target $57.0B, eCPS $59.3B, MP $356.9B

Desired direction

Fix these with source/model/target-support improvements, not by relaxing the sound comparison gate or treating eCPS as ground truth. Use eCPS only as a benchmark/control. The model should explain why Microplex is placing weight on these surfaces and whether the support comes from source data, donor imputation, formula-derived outputs, or calibration overfit.

Start with:

1. Add sidecar diagnostics that tie worst target rows to source-support/role surfaces: filing status, AGI bin, SSI recipient/value source, Medicare Part B premium support, ACA state support.
2. Diagnose whether the HoH AGI-bin failures are from remaining tax-unit role support, filing-status formula/output mismatch, AGI construction, or optimizer weight concentration.
3. Diagnose whether SSI/Medicare/ACA failures are exported input issues, formula-derived benchmark targets, donor-source support issues, or calibration target omissions.
4. Rebuild or rematerialize the small ASEC+ACS100k candidate after each fix and rerun the sound matched-N symmetric eCPS comparison.

Acceptance criteria

• The sound comparison artifact includes the target diagnostics sidecar added in Add target diagnostics to eCPS comparison #85.
• Top regressions no longer dominated by HoH AGI-bin count/AGI cells, SSI, Medicare Part B premiums, and ACA spending without a clear source-support explanation.
• Candidate improves on full and holdout loss versus the current-main 3.3934 / 0.5038 baseline under matched-N symmetric refit.
• Protected-family and core-family floors in the mp-300k gate remain active; no gate relaxation is used to pass.

[MaxGhenis]

Merged PR #87: #87

This adds the support sidecar needed to diagnose this issue rather than only seeing target-loss symptoms.

What is now captured in support_audit.json and the comparison summary:

• Head-of-Household AGI-bin support for the bins currently dominating IRS regressions (20k-25k, 25k-30k, 30k-40k, 200k-500k, 500k-1m, 1m+)
• SSI recipients and SSI dollars by age group (all, under_18, 18_to_64, 65_plus)
• Medicare Part B premium support by age bucket, including the young-age buckets that are currently bad targets
• state ACA PTC spending and positive-household support
• existing critical input, filing status, MFS high-AGI, marketplace enrollment, and state-age support diagnostics

Real validation artifact from current matched/refit H5s:

/Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/sound_ecps_replacement_comparison/support_audit_with_issue86_surfaces_20260529.json

The new audit confirms the visible failure mode with formula/output support, not just loss labels. Examples:

• HoH AGI 25k-30k: candidate weighted count 2.79M vs eCPS 0.27M, delta +2.52M; candidate weighted AGI $76.0B vs eCPS $7.5B.
• HoH AGI 500k-1m: candidate weighted count 0.86M vs eCPS 0.04M, delta +0.82M; candidate weighted AGI $564.7B vs eCPS $30.9B.
• SSI age 65+: candidate recipients 32.13M vs eCPS 3.81M, delta +28.32M; candidate SSI $279.0B vs eCPS $27.3B.
• Medicare Part B age 10-19: eCPS has zero support; candidate has positive support, matching the current young-age target regression.
• ACA PTC top state gap in the real audit is TX: candidate $31.8B vs eCPS $8.1B, delta +23.7B.

CI passed before merge (implementation-tests, site-snapshot, build-docs; artifact-gates skipped as expected). Next work should use this sidecar to determine whether the HoH failures come from tax-unit role support, filing-status formula/storage, AGI construction, or weight concentration, and whether SSI/Medicare/ACA failures are source-support or calibration-target omissions.

[MaxGhenis]

Opened PR #88 to address the HoH filing-status support issue surfaced by the new audit: https://github.com/PolicyEngine/microplex-us/pull/88\n\nThe fix stops converting role-flag tax units to HEAD_OF_HOUSEHOLD solely because they have dependents. Explicit HoH evidence from filing_status_code == 4 is still preserved. Local checks passed: ruff on us.py/test_us.py and the full tests/pipelines/test_us.py suite.

[MaxGhenis]

Follow-up after PR #88 merged and rematerializing the existing 100k-household candidate from the same calibrated source under current main.

Artifacts:

• Rematerialized H5: /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/materialized_after_hoh_status_fix_20260529/policyengine_us.h5
• Raw support audit vs latest local eCPS: /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/materialized_after_hoh_status_fix_20260529/support_audit_vs_latest_ecps_20260529.json
• Sound matched/refit comparison: /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/sound_ecps_replacement_comparison_after_hoh_status_fix_20260529/sound_ecps_replacement_comparison.json
• Target diagnostics: /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/sound_ecps_replacement_comparison_after_hoh_status_fix_20260529/target_loss_diagnostics.json
• Support audit sidecar: /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/sound_ecps_replacement_comparison_after_hoh_status_fix_20260529/support_audit.json

Effect of the HoH filing-status fix:

• Full candidate tax-unit HoH count dropped from 35,873 to 19,032, with entity counts otherwise unchanged: 100,000 households, 245,714 persons, 130,980 tax units.
• Raw weighted HoH support vs latest eCPS dropped from 53.63M to 28.76M; eCPS is 17.79M.
• Sound matched/refit PE-native broad loss improved materially but is still not close to eCPS: candidate 1.9704 vs eCPS 0.1646. Previous current-main candidate loss was 3.3934, so this was a real improvement, not a full resolution.

Remaining top blockers in the new sound comparison:

• Medicare Part B premiums by young ages are now the largest single target regression, e.g. nation/census/medicare_part_b_premiums/age_10_to_19.
• SSI remains far too high, especially 65+: candidate weighted recipients 32.1M vs eCPS 3.8M in the support audit.
• HoH AGI support is still high after Avoid inferred HoH filing status from role dependents #88: matched/refit HoH weighted count 29.0M vs eCPS 17.8M, with top gaps in 30k-40k, 25k-30k, and high-AGI HoH bins.
• ACA PTC/state ACA spending remains high, with TX and CA still the largest dollar gaps.

Next likely production fixes should target Medicare Part B export/support and SSI take-up/value assignment before another full replacement claim.

[MaxGhenis]

Opened PR #89 for the next #86 blocker: https://github.com/PolicyEngine/microplex-us/pull/89\n\nIt gates on in PE person-input augmentation. This removes logically impossible Part B support already present in the current source artifact, including all under-65/no-Medicare premiums, before export/scoring.

[MaxGhenis]

Correction to the prior comment: PR #89 gates medicare_part_b_premiums on has_medicare in PE person-input augmentation.

[MaxGhenis]

PR #89 merged: #89

Local verification before merge:

• ruff check src/microplex_us/pipelines/us.py tests/pipelines/test_us.py: passed
• focused test_us.py -k "part_b or augment_policyengine_person_inputs": 5 passed
• full tests/pipelines/test_us.py: 148 passed

On the current 100k-household source artifact, the new guard changes Part B support before export as follows:

• total weighted Part B premiums: $125.63B -> $120.60B
• under-65/no-Medicare weighted Part B premiums: $1.92B -> $0
• 65+ weighted Part B premiums: $119.19B -> $116.08B

This should reduce the impossible young-age Part B target regressions, but it will not fully solve #86 by itself because under-65 people with has_medicare still carry about $4.52B of Part B premiums against much smaller young-age targets. Next likely fix is either age-bucket/calibration-aware Part B rescaling or a donor-imputation constraint keyed on Medicare eligibility/age/disability.

[MaxGhenis]

Opened PR #90 for the SSI support/export-contract part of this issue: #90

Evidence from the current mp-300k artifact:

• Source/materialized person parquet has plausible reported SSI support: 6.07M weighted positive SSI records and $63.6B reported SSI.
• The exported H5 currently drops takes_up_ssi_if_eligible, so PE-US uses the variable default (True) and calculates SSI for every formula-eligible person. That yields $358.5B SSI and 40.8M recipients in the current Microplex artifact.
• Latest eCPS, by comparison, calculates $59.3B SSI and 7.69M recipients.

PR #90 does not export ssi or ssi_reported as direct PE inputs. It only uses explicit/source-reported SSI support to seed takes_up_ssi_if_eligible, and allows PE takeup input variables through the safe export surface.

Diagnostic artifact: /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_family_relationship_units_20260529/materialized_after_hoh_status_fix_20260529/ssi_takeup_export_probe_20260529.json.

Important caveat: source-positive SSI takeup alone is not the final SSI model. A probe with takes_up_ssi_if_eligible=(source ssi > 0) lowers calculated SSI to $11.5B and 1.25M recipients because source-reported recipients and PE-US formula eligibility do not overlap enough. So #90 fixes the accidental 100% takeup export bug and preserves the right contract, but the next step is a data-oriented SSI takeup assignment calibrated against SSA/Arch recipient and amount targets by aged/disabled/blind domains.

[MaxGhenis]

Merged PR #90: #90

Merge commit: b3d648b1eb2d669c828950dc73caac2d0d402f7b.

This fixes the immediate export-contract bug: Microplex can now carry PE takeup inputs, derives takes_up_ssi_if_eligible from explicit/source-reported SSI support, and still blocks ssi/ssi_reported as direct PE output overrides.

Remaining SSI work for this issue: build a calibrated/data-oriented SSI takeup assignment rather than relying only on source-positive SSI support. The probe shows why: source-positive takeup avoids the $358.5B/40.8M accidental full-takeup result, but because source SSI recipients and PE-US formula eligibility overlap poorly, it lands around $11.5B/1.25M recipients rather than the latest eCPS $59.3B/7.69M.

[MaxGhenis]

Opened PR #91 to make the SSI takeup assignment data-oriented without exporting source SSI as a model output: #91

The patch sets takes_up_ssi_if_eligible by matching reported source SSI dollars within aged and under-65 groups against PolicyEngine full-takeup SSI potential. Probe on the current mp-300k artifact: all-eligible takeup was ~$358.5B SSI / 40.8M recipients; source-positive-only takeup was $11.5B / 1.25M recipients; age-amount calibrated takeup is $63.6B / 7.06M recipients, close to source reported dollars ($63.6B) and latest eCPS total ($59.3B). Blind/aged mix still needs separate follow-up.

[MaxGhenis]

Opened PR #92 for the next SSI support gap: #92

Current mp-300k materialized H5 lacks is_blind, while latest eCPS exports it. The source persons parquet has difficulty_seeing support covering 4,017 rows / about 5.55M weighted people, so PR #92 derives is_blind from that proxy and adds it to the safe PE export surface. This should prevent blind SSI cells from being forced to zero on the next rebuild.

[MaxGhenis]

Merged the two SSI follow-ups:

• Calibrate SSI takeup from reported support #91 Calibrate SSI takeup from reported support merged at b30aad0. This replaces all-eligible/source-positive-only SSI takeup with age-group amount calibration against PE full-takeup SSI potential.
• Export blind status from visual difficulty #92 Export blind status from visual difficulty merged at 1616016. This exports is_blind from the difficulty_seeing proxy and adds it to the safe PE export surface.

Next validation step is a rebuilt/smoke MP artifact from current main to confirm SSI totals and blind-domain support improve in the actual replacement comparison, then continue remaining support blockers such as ACA PTC state gaps and HoH AGI.

[MaxGhenis]

Validation smoke after #91/#92 is in /Users/maxghenis/CosilicoAI/microplex-us/artifacts/ssi_blind_smoke_after_pr91_pr92_20260529/ssi_blind_smoke_summary.json and ssi_category_target_definition_smoke.json. On the existing mp-300k artifact, the merged changes move PE-calculated SSI from ~$358.5B / 40.8M recipients to ~$63.6B / 7.0M recipients, and blind support is no longer zero.

Opened #93 for the next definition issue: #93. Arch SSA ssi_category facts are qualification-category targets, so the adapter should preserve them as ssi_category == AGED/BLIND/DISABLED rather than map them to broad proxy booleans like is_blind or age 65+.

[MaxGhenis]

Updated validation after the current export-contract fixes and the tenure enum fix in #104.

New artifact:

/Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_current_main_after_export_contract_20260529/sound_ecps_replacement_comparison_after_tenure_fix_clean_usdata_env_20260529/sound_ecps_replacement_comparison.json

Sidecars:

• /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_current_main_after_export_contract_20260529/sound_ecps_replacement_comparison_after_tenure_fix_clean_usdata_env_20260529/target_loss_diagnostics.json
• /Users/maxghenis/CosilicoAI/microplex-us/artifacts/small_asec_acs100k_current_main_after_export_contract_20260529/sound_ecps_replacement_comparison_after_tenure_fix_clean_usdata_env_20260529/support_audit.json

Run contract:

• matched household count: 41,314 vs 41,314
• symmetric refit: true
• score-candidate-only: false
• holdout target fraction: 0.2 (738 holdout targets)
• eCPS refit recovery: passed
• refit objective matches scoring: true
• target count kept: 3,702 (1,220 national, 2,482 state)

Result:

• Microplex refit PE-native broad loss: 1.7408481701051062
• eCPS refit PE-native broad loss: 0.7685318623404788
• delta: +0.9723163077646274; candidate does not beat baseline
• Microplex initial -> refit: 1.902312877799414 -> 1.7408481701051062
• eCPS initial -> refit: 0.7689706412507552 -> 0.7685318623404788
• train losses: MP 1.460624649665914, eCPS 0.7433881085366006
• holdout losses: MP 0.28022353265381306, eCPS 0.025143700961373777
• target wins: MP 646, eCPS 2,977, ties 79

This is a large improvement from the prior current-main artifact in the issue body (3.3934 MP loss), but still not replacement-ready. Main loss deltas are still concentrated in:

• national_irs_other: +0.7391
• state_agi_distribution: +0.1592
• state_aca_enrollment: +0.0779
• state_age_distribution: +0.0124

Specific support audit signals still worth prioritizing:

• HoH AGI/count bins remain top regressions, including 20k-25k, 25k-30k, 500k-1m, and 1m+ taxable HoH cells.
• ACA/marketplace support is too high in several large states; e.g. CA candidate ACA PTC is $32.4B vs eCPS $7.9B, TX $32.2B vs $16.8B.
• SSI is now under, not over: total candidate SSI $1.48B / 156k recipients vs eCPS $30.1B / 4.14M, with the largest gap at age 65+.
• SNAP is one area where MP improves over eCPS in this run: state SNAP cost and household families have negative loss deltas.
• Protected family check: MP slightly beats eCPS on self-employment and disability, but loses on capital gains, dividends, interest, wages, SSI, retirement income, and household net income.

Notes on provenance: the scoring used a clean policyengine-us-data upstream-main worktree at 480de9930fd9fa0e97d165155f6d4961a4c2ec83 for the target/loss code path. The baseline H5 was the local enhanced_cps_2024.h5 artifact from the existing PE-US-data checkout, not a newly rebuilt H5 from the clean worktree.