[MaxGhenis]

Summary

PolicyEngine-US rules that key on actual SSI receipt read (ssi > 0) | receives_ssi. Examples are state supplements, the Medicaid SSI categories, SNAP categorical eligibility, Head Start, and county GA exclusions. Colorado's AND-CS joins them once PolicyEngine/policyengine-us#10145 merges. No US release carries receives_ssi, so every one of these rules depends only on modeled SSI. This PR makes the base build carry the survey fact, and makes the release coverage gate require it.

Draft until Max rules on the methodology question below (d1675). Merging changes what the next national default carries. Nothing here touches a published release.

What changes

• Producer. cps_carried.reported_ssi_receipt marks a person when SSI_VAL > 0 and at least one reported reason is their own (or none is recorded).

  • Respondents whose only RESNSSI reasons are "on behalf of a disabled/blind child" (codes 3 and 4) are payees, not recipients, so they are not marked. That is 133 rows and 0.45M weighted people on the pinned default.
  • ASEC asks SSI only of people aged 15 and older, so no child under 15 is marked.
  • The producer refuses positive amounts when the reason columns are missing.
  • The annual report is broadcast to all months, like receives_snap, receives_tanf and receives_wic.
  • No amount input is filled. Sending the reported amount as the ssi input would replace the federal formula for every reporter, so SSI reforms would stop moving them. ssi_reported is deprecated and feeds nothing.

• Release gate. receives_ssi is a required coverage input with no reviewed exclusion (US_ASEC_REPORTED_SSI_RECEIPT_REQUIRED_INPUTS). us_release_input_coverage_gate also checks two invariants:

  • every marked person reports a positive amount of their own: SSI_VAL on CPS rows, native ssi_reported on stacked ACS rows;
  • every marked CPS-measured person carries takes_up_ssi_if_eligible. The take-up stage anchors CPS reporters only, so ACS rows are counted, not judged.

  Release preflight treats these row-dependent failures as at risk on the L0 path and as certain on the full pool.

• ACS rows (multispine pool and ACS local chain). map_acs_native_inputs maps receives_ssi natively from a positive SSIP (blank below age 15 is no receipt). The pool lists it as complete on both spines (_POOL_NATIVE_COMPLETE_OUTPUTS), so it is never a QRF transfer target, and the stacked surface stays at 134 targets.

  • Only the pool's engine-input provisioning (assembled native 5→6) and the checkpoint identity move. They were re-derived with tools/repin_us_pool_engine_contracts.py and tools/spec_engine_coverage.py.
  • The release coverage manifest was regenerated with tools/build_us_release_input_coverage_manifest.py.
  • ACS records no reason for receipt, so ACS rows cannot exclude child payees.

• Docs. docs/us-reported-ssi-receipt.md, plus measured evidence in docs/evidence/us-reported-ssi-receipt/ (aggregates only).

Invariants (property-tested with Hypothesis)

• A marked person always reports a positive SSI_VAL.
• A person whose recorded reasons, in either position, are all on-behalf-of-child codes is never marked.
• Every other positive reporter is marked.
• Adding an own reason never unmarks a person.
• The derivation is idempotent and keeps an existing column.
• The consistency check fails exactly when a marked person lacks a positive observed SSI_VAL or the take-up anchor.
• Round trip: a receipt derived from SSI_VAL, with the take-up stage's anchor, always passes the check.

Measured effect (pinned default populace_us_2024.h5, 2026; fixed-weight sensitivity)

Method and full tables are in docs/evidence/us-reported-ssi-receipt/README.md.

• Approach: PolicyEngine-US main fab18b7d6b and #10145 f450d4ab00, run on every household with a reporter plus every Colorado household. receives_ssi is a person input with no cross-household interaction, so these deltas are the national deltas. The run checks this on Colorado's non-reporter households, whose deltas are exactly zero.
• No full-population run. A full-population rerun is held back by the host reservation until 10/21.

Change on today's default (fixed weights, not recalibrated; no ACS rows)	With receives_ssi
Federal SSI	$0
Household net income	+$339M
SNAP	+$261M
Early Head Start	+$1,156M
ACA PTC	−$863M
Medicaid enrollees	+1.41M people
State supplements (MI, WA, WY)	+$14M
Colorado AND-CS under #10145	$0 (unchanged)

Nearly all of this comes from reporters whom the model pays $0. They are treated as SSI recipients for categorical purposes while their modeled SSI income stays $0. Of 5.15M own-reason reporters, the model pays 2.79M. The other 2.35M get nothing for these reasons:

• 1.03M PUF clones without the disability criterion. PUF tax-detail clones of under-65 reporters do not inherit the ASEC meets_ssi_disability_criteria anchor, although they carry SSI_VAL and the take-up anchor. Anchoring them would add $6.2B of federal SSI before take-up recalibration.
• 0.78M income-test failures. On PUF clones these have a median countable income of $5,265/month from PUF-imputed incomes; on ASEC rows, $1,750.
• 0.45M resource-test failures. Zeroing every reporter household's liquid assets brings only 0.16M of them into payment (cf. Imputed savings put CPS SSI reporters over the resource limit, so reforms 'add' people who already report SSI #1184).

Main's $233M Colorado supplement goes to 11 records, none of which reports SSI. They are modeled-eligible non-claimants, so reported receipt restores none of it.

Methodology choice for Max (d1675)

More than one reading is defensible, so this is queued as a decision. There are two separate axes.

A. CPS rows (the national default):

1. As implemented. Receipt goes on ASEC rows and their PUF clones, the same as the take-up anchor and the other receipt inputs.
2. ASEC rows only. Clones carry no receipt, which matches the disability-criteria anchor. This is about 40% of the effect (+0.37M Medicaid enrollees, +$141M net income).
3. Hold until the attribute fixes land. These are SSI anchors treat adults who receive SSI only on behalf of a child as disabled SSI claimants #1313 (payees), PUF clones of under-65 SSI reporters keep the take-up anchor but not the SSI disability-criteria anchor #1314 (PUF-clone disability anchor) and Imputed savings put CPS SSI reporters over the resource limit, so reforms 'add' people who already report SSI #1184; then (1).
4. Don't carry receipt. Keep receipt modeled-only.

B. ACS rows (stacked and local-area releases only; the national default has none):

1. As implemented. Receipt is mapped natively from SSIP. ACS cannot exclude child payees, and ACS reporters are not take-up-anchored until SSI take-up anchors CPS SSI reporters but not ACS SSIP reporters on stacked frames #1317, so some carry receipt with $0 modeled SSI. This effect is unmeasured.
2. No ACS receipt until SSI take-up anchors CPS SSI reporters but not ACS SSIP reporters on stacked frames #1317 anchors ACS take-up.

Recommendation: A1 + B1. The receipt is a leaf survey fact that the raw-only doctrine carries. The disagreements come from imputed attributes, fixed where they arise (#1313–#1315, #1317).

Review

Round 1 (GPT-6.1 Sol, dd6e51791) asked for changes; all five findings are addressed in 1466d4195:

1. The QRF-transferred ACS receipt could break the take-up invariant. The take-up arm is now scoped to CPS-measured rows.
2. ACS has native SSIP. ACS receipt is now native and never transferred.
3. Dry-run classification: receipt-consistency failures are at risk on L0, with their counts kept.
4. The payee rule is now order-free, with an independent oracle.
5. The evidence claim about the two engine trees is corrected.

Round 2 (Sol, 1466d4195) found all five round-1 findings resolved and asked for six smaller changes. They are addressed in 902c6d2d8:

• ACS options are stated above.
• An ACS receipt-without-take-up counter is added, and SSI take-up anchors CPS SSI reporters but not ACS SSIP reporters on stacked frames #1317 is filed.
• A 15+ blank SSIP is now refused in the mapper.
• A receipt with no amount column now fails.
• Stale text is fixed.
• Tests now cover the no-transfer property.

Verification

• ruff check . passes. python3 tools/ci_test_plan.py verify passes.
• Focused suites pass: test_us_reported_ssi_receipt (new), test_us_stacked_spine, test_us_spine_blindness, test_us_puf_support, test_us_acs_transfer, test_us_plan, test_release_input_coverage, test_spec_engine_typed_closure.
• Full engine-free and engine-us categories run in CI (a host reservation holds full local suites).
• tools/generate_us_bundle_from_constants.py --check and --engine-lock-only --check pass. tools/repin_us_pool_engine_contracts.py reports current.

Follow-ups filed

• SSI anchors treat adults who receive SSI only on behalf of a child as disabled SSI claimants #1313: SSI anchors (take-up, disability criteria, is_disabled) treat on-behalf-of-child payees as disabled SSI claimants. The model pays 0.16M of them $1.5B.
• PUF clones of under-65 SSI reporters keep the take-up anchor but not the SSI disability-criteria anchor #1314: PUF clones of under-65 reporters keep the take-up anchor but not the disability-criteria anchor (1.03M reporters).
• PUF clones of SSI reporters get imputed incomes far above the SSI income limit #1315: PUF-imputed incomes on clones of SSI reporters are far above the SSI limit (0.32M).
• Imputed savings put CPS SSI reporters over the resource limit, so reforms 'add' people who already report SSI #1184: comment with updated resource-test numbers.
• SSI take-up anchors CPS SSI reporters but not ACS SSIP reporters on stacked frames #1317: the SSI take-up stage does not anchor ACS SSIP reporters on stacked frames.

axiom: n/a: Microcosm data input, no rules change.

🤖 Generated with Claude Code