Plan: start the projections from Microcosm's US population (docs; draft #42 registration, not posted)

[MaxGhenis]

What this adds

Docs only. Nothing here runs.

• docs/plans/microcosm-start.md is a plan for starting the Social Security projections from Microcosm's US cross-section, the population PolicyEngine-US runs on, instead of the 2011 PSID cohort.
• docs/plans/microcosm-start-registration-draft.md is a draft issue Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 registration of the tests that would score that start: two gates and two report-only checks. It is not posted. Posting waits for Max's go (d947).
• docs/plans/microcosm-start-evidence/ holds a structure-only script, and its output, for three cached Microcosm releases. It records column lists, source years, support channels and effective sample sizes from household weights. No outcome variable was summarized. The output reproduces gate w1's recorded effective size of 14,327.8 exactly.

The plan in brief

• Earnings histories. Each frame person gets a history built backward from their start-year earnings rank, with PSID donors (rank-kNN, gate 1's passing method).
  • QRF chains failed gate 1 in both directions: the two-predictor chain forgets (ten-year autocorrelation 0.31–0.37 against 0.54), and versions given a fixed person effect overshoot (0.62–0.67).
  • Three candidates are frozen in the registration, run once together, and adopted by a registered rule: rank-kNN, then rank-kNN with marital strata, then a QRF with the predictors H1 can align across the PSID and the frame.
  • Years the PSID lacks use gate_epuf_fill's certified fills, so H1's run waits for that gate's verdict (d927).
• Starting stock. People already receiving benefits keep the benefit the frame reports, as the PSID projection does. PIAs are inferred for retired and disabled workers, with named limits.
• Income and wealth. Taken from Microcosm and mapped to the stage that produces each item: the CPS, the 2015 PUF, the 2022 SCF and the 2023 SIPP. They age by growth plus rank within cell. A1 reports how well that works against PSID panels.
• Gates.
  • H1 scores the deployed history procedure. One arm uses held-out EPUF careers on a new salted split; the other uses held-out PSID families.
  • Gate 3 gets its first cells: one-year growth against SSA, locked separately once the engine exists.
  • W2 (report-only) checks the starting stock against SSA's Supplement tables.
• Order and effort. 70–109 agent-days and 22–29.5 review days, both assumptions. The history work can start before funding.
• Schedule. The plan is reconciled with the NASI 24-month PlanGraph draft, whose months 5–9 are built on the PSID projections and which leaves the Microcosm port out of the 24 months.
• Risks.
  • Weight concentration: effective size about 14,300 (W1 frame), about 27,000 (Registration 19's) and about 180,000 (a dense build that lacks the needed columns).
  • No education column.
  • 85.6 percent of older beneficiaries are married.
  • A 2024 frame cannot run the 2010 DYNASIM3 tests.
  • PolicyEngine-US takes Social Security benefits as inputs.

Independent review (Opus 5.5 lanes on Subfleet)

Round	Head	Verdict	Report	Response
1	24427335	CHANGES REQUIRED (3 blocking, 16 major, 19 minor)	review-r1.md	response-r1.md
2	26d42793	CHANGES REQUIRED (6 major, 14 minor)	review-r2.md	response-r2.md
3	b81ff1cd	APPROVE AFTER LISTED FIXES	review-r3.md	response-r3.md
Delta 1	fd4a8ec4	APPROVE	review-r3-delta.md	three optional follow-ups applied
Delta 2	bb8bacf1	APPROVE	review-r4-delta.md	—

All reports are in ~/reviews/microcosm-start-plan-20261004/review/. Main changes the reviews forced:

• the opening-stock statement corrected;
• candidates frozen together with an adoption rule;
• a fresh EPUF split, with H1 gating the deployed procedure;
• gate 3's growth statistic and a runnable bite check;
• the frame's source_year read as the income year;
• the dense-build column gap;
• the forward-law hand-off read as u_w, not the two-year lag.

Decisions

The go to post the registration is queued as d947. The plan's "Decisions needed" also lists:

• which frame for dynamics;
• the tolerance multipliers at lock;
• the 2010 question.

🤖 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 5, 2026 3:01am UTC

[MaxGhenis]

Review round 1: CHANGES REQUIRED, answered.

The independent review is an Opus 5.5 Subfleet lane, at head 24427335. It found 3 blocking, 16 major and 19 minor findings, and confirmed most of the numbers.

The main fixes, at head 26d42793:

• H1's candidates are frozen in the registration, run once together, and adopted by a registered rule (H-A, else H-B, else H-C).
• H1's SSA arm has its own salted EPUF split instead of reusing gate_epuf_fill's TEST.
• Gate 3 scores the signed growth gap against SSA. Its partition and constants are fixed from the start-year frame before any run, and it has a bite check.
• Opening stock. The statement was backwards and is corrected: the PSID projection already carries recorded amounts forward (cola_track_a/benefits.py:747-790).
• The frame's source_year is the income year (2022–2024, from the 2023–2025 supplements).
• The dense Microcosm build lacks the columns the plan needs. Decision 2 is reworked, and the evidence script now lists household columns.

The point-by-point response is in the review record (~/reviews/microcosm-start-plan-20261004/review/response-r1.md). Round 2, a check of each fix, is running. The PR stays a draft until it agrees.

[MaxGhenis]

Review round 2: CHANGES REQUIRED, answered at b81ff1cd. Round 2 confirmed most of the round-1 fixes. It raised six major points, all now fixed. N1: H1's SSA arm now gates the deployed procedure (PSID donors, certified fills refit on H1's training part, conversion to capped wage-base shares), with EPUF donors as a report-only configuration. N2: gate 3's bite check runs on an invented frame, and gate 3 locks separately once its engine exists. N3: the forward-law claim is corrected; the law reads start-year earnings and u_w. N4: P1 is defined on the frame's own concept. N5: the PSID arm's bite check uses a reserved bite part. N6: arm E's partition has a 5 percent false-fail budget. The point-by-point response is in ~/reviews/microcosm-start-plan-20261004/review/response-r2.md. A short round-3 confirmation is running.

[MaxGhenis]

Merge record. Merged at head bb8bacf1 as 03a7d231 (merge commit, pinned with --match-head-commit). Checks at that head: gh pr checks exited 0 (lint, build, pytest (3.14) and every pytest shard passed). Mergeable: CLEAN. No changes-requested review. Independent review: Opus 5.5 lanes on Subfleet, three rounds plus two delta checks, the last of which approved this head; reports are in ~/reviews/microcosm-start-plan-20261004/review/. This merges a plan and a draft registration only. Posting the registration on issue #42 waits for Max's go (d947).