[MaxGhenis]

Registration 19: specification, code, and the released one-shot results. The registered run took place on 2026-10-06 at the registered commit 3ba0d020 (tag registration-19-sa1). Max released the results on d905 (2026-10-10), and they were posted on issue #42 here.

Warning

Frame defect: microcosm#1178. The frame's PUF-support channel erases CPS-reported Social Security for 11.8 million weighted people, 9.2 million of them aged 65 and over. Only 15.8% of the model's SSI recipients aged 65 and over have Social Security, against SSA's 57%. The results are the registered result on that frame, published as the registration commits; they are not an estimate for the US population. A rerun on the corrected frame will be registered separately once #1178 is fixed.

Merge with a merge commit, never squash or rebase: the registered commit 3ba0d020 and the results commit 5fc28d08 must stay in history.

Results in this PR

• runs/pe_us_depletion_cut_population_v1.json and its .env.json sidecar: the bytes whose SHA-256 (ef9366be…, 360a3b99…) were posted on issue Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 before release (comment 6014920955). A pin test (tests/bridge/test_pe_us_depletion_cut_population_v1_artifact.py) holds them.
• docs/analysis/pe_us_depletion_cut_population_20261001/: the report and the chart.
• Run: Modal call fc-01M48CG9VMP741AWN9A163FE9B, first attempt, exit 0. Every refusal and row-validity check passed. An independent recomputation from the raw arrays reproduced all 12,319 compared values.

What this is

PR #506 showed that for an illustrative low earner with $1,500 in the bank, SSI replaces a depletion-sized Social Security cut dollar for dollar. At the NASI meeting on 2026-10-01, participants asked how often that holds for real beneficiaries, since SSI's asset limit excludes many people, and asked for results by age, sex, marital status, race and income quintile.

This PR adds the registered analysis that answers it: the same cut (22 percent OASI, with 17 percent OASDI as the sensitivity) on Microcosm's populace-us-2024 frame, which carries imputed bank, stock and bond assets, through PolicyEngine-US 2.18.0, with and without SSI's resource test.

• Specification (ratified, frozen): docs/design/pe_us_depletion_cut_population.md, sa1-ratified-1, SHA-256 4672d084aaf18d492eff30ca6961d511926086c8f73fc60368849b3589e11e1f.
• Registration: Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 (comment) (first of two comments).
• Labels on every output: frame-relative; 2026 law and prices, not a projection; static; registered one-shot, not blind.

What it adds

• src/populace_dynamics/bridge/depletion_cut_population.py: the registered arithmetic on arrays. The cut, the marital-unit replacement labels (full, part, none), the resource tests for rows R0, R3 and R4, the cells, the statistics and the household bootstrap.
• src/populace_dynamics/bridge/depletion_cut_population_child.py: the worker that runs inside the PolicyEngine-US interpreter. It always passes an explicit dataset, refuses the real frame outside a registered run, and removes the resource test with a structural reform.
• scripts/run_pe_us_depletion_cut_population_registered.py: the one-shot entry point, with the preflight refusals, the run-time checks and the labelled outputs.
• Tests: unit and property tests, a specification test that holds the code's constants to the spec's machine-readable block, script tests, and live oracle tests on invented households.

It also fixes one thing in #506: its artifact test ran git diff --quiet <commit> HEAD -- src scripts, so it would fail as soon as any later change touched src/ or scripts/. The diff now covers only the files that artifact records as its dependencies.

Invariants

Stated in the specification and tested as properties:

• The vectorized cut equals PolicyEngine-US bridge: depletion-cut offsets for illustrative households #506's payable_monthly_benefit element by element; 12m <= a < 12(m + 1) holds exactly; a payable share of 1 changes nothing; the OASI cut never touches the disability component.
• The three replacement labels partition the cut beneficiaries, and their shares sum to one.
• Person by person, SSI with an asset test equals SSI without it times the pass indicator (checked at run time; rows R3 and R4 rest on it).
• Household pooling (R4) blocks at least as often as spousal deeming (R3).
• One-way cells partition each universe; quintile cutpoints follow the registered weighted rule, and ties share a quintile.
• The bootstrap is deterministic under its seed and does not depend on the order of people within a household.

What the review found

Four independent referees reviewed the specification in two rounds, plus two re-checks before ratification; §20 of the specification and the registration comment list them. Findings that changed the design:

• PolicyEngine-US 2.18.0 does not deem a spouse's resources (20 CFR 416.1202(a)); row R3 applies that rule.
• It caps a deemed benefit at the individual rate only, where 20 CFR 416.1163(e)(2) also subtracts the person's own countable income; the "none" label therefore includes SSI that falls.
• The frame puts each household's assets on one person, so rows R3 and R4 show how the result depends on that.
• The frame's beneficiaries aged 62 and over are 85.6 percent married, against about 60 percent in published figures. Every output carries that as a named difference.

Tests

The code was tested on Modal (Linux, the frozen environments of #506's runner) with invented data only: 239 tests at the registered commit, plus a frame-sized invented dry run and control-path rehearsals of the runner. CI runs the full suite here, including the tier census: unit 6,166; artifact 3,540; integration_psid 1,341; reproduction_legacy 520; oracle_policyengine 232.

axiom: n/a: analysis code, no policy change. (Two PolicyEngine-US SSI encoding gaps found here are filed separately.)

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
social-security-model	Ready	Preview	Oct 12, 2026 9:15am UTC

[MaxGhenis]

Merge audit. Merged at 2026-10-12T11:54:00Z as merge commit 42bd11a1, pinned to the reviewed head e5fc2165 (--match-head-commit). No squash or rebase.

Gates at merge

• gh pr checks 511 exited 0; all 11 checks passed (lint, build, pytest, pytest shards 1–4, graph-integration on 3.13 and 3.14, Vercel). The Vercel preview first failed before building, with no log; a redeploy of the same commit passed.
• mergeable was MERGEABLE, the PR was not a draft, and there was no CHANGES_REQUESTED review.
• Independent reviews by GPT-6.1 Sol, all APPROVE: at 9e0c814f (code and specification), at 7615fc15 (the released results and a master merge), and at e5fc2165 (the final master merge and tier counts).
• The release of the results was Max's call, made on d905 (2026-10-10). The results comment is on issue Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42: Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 (comment). It discloses microcosm#1178 first.

Verified on master after the merge

• The registered commit 3ba0d020 (tag registration-19-sa1) and the results commit 5fc28d08 are both in master's history.
• runs/pe_us_depletion_cut_population_v1.json hashes to ef9366be…, the value posted on Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 before release.
• Tier counts: unit 6,550; artifact 3,929; integration_psid 1,346; reproduction_legacy 520; oracle_policyengine 232. These are master's counts plus this PR's additions, confirmed by CI's tier census.

Still open

• Registration 20, the rerun on a frame without the defect, stays an unregistered draft until microcosm#1178 is fixed (PR microcosm#1186) and the frame is recertified.
• No numbers go to outside readers until then.