Replication anchors: DYNASIM-scored reforms and the minimum model surface for each

Replication anchors: DYNASIM-scored reforms and the minimum model surface for each

The gate-1 program certified an earnings-history generator; the reform-delta diagnostic (#72) showed it reproduces aggregate reform incidence within noise on formula-side reforms. The natural next question is external: can this stack replicate reform analyses that DYNASIM has published? This issue catalogs the published anchors (every number transcribed from a fetched Urban Institute PDF, 2026-07-07), maps each provision to the minimum model components it needs, and proposes the phased build. It is a design issue, not a registration; each phase would lock its own pre-registered gate before its first scored run, exactly as gate 1 did.

The anchor shortlist (ranked by table quality × definition precision × model surface)

Cross-model triangulation: SSA OCACT single-provision scores (ssa.gov/oact/solvency/provisions, HTML tables per run) and bill memos (Social Security 2100 baseline −2.84% payroll under the 2018 TR), and SSA MINT Policy Brief 2009-01 on taxable-maximum distributional effects. SSA pages 403 on programmatic fetch; verify manually before publication.

Component requirements per provision

Components: EH earnings histories (certified, gate 1) · SF statutory PIA formula (pinned oracle, cross-validated) · CA claiming ages · M differential mortality · MS marriage/survivorship · FC fertility/caregiving · DI disability · REV payroll/revenue side · LR long-run projection under Trustees assumptions.

(✓) = needed only for poverty/household or subgroup rows, not the core benefit calculation.

Data already staged for the demographic components

All on disk at ~/PolicyEngine/psid-data, label-verified: mh85_23 (marriage histories), cah85_23 (childbirth/adoption histories), MX23REL (person-relationship map), death years in ind2023er — exactly the holdout basis gates.yaml gate 2 already names. External estimation/validation sources: SSA cohort life tables (differential mortality), Statistical Supplement claim-age distributions (claiming), Trustees intermediate assumptions (projection).

Proposed phases (each behind its own pre-registered gate)

• Phase A — formula mechanics on observed careers (buildable now). Replicate the incidence structure of anchor 1's PI/PPI tables (%-of-scheduled by quintile) and OCACT-style single-provision patterns on completed PSID careers, with the population-concept delta (observed cohorts vs DYNASIM's 2050 retirees) documented rather than hidden.
• Phase B — claiming + differential mortality. Unlocks the COLA rows, NRA→70, the 80–85 age rows, long-term bonus. Gate: mortality hazards vs SSA cohort life tables + PSID deaths; claim-age distribution vs the Statistical Supplement.
• Phase C — marriage/survivorship (gate 2 proper). Unlocks anchor 5 (the survivor/earnings-sharing winners/losers tables) — DYNASIM's signature territory. Holdouts: the staged mh85_23/MX23REL files.
• Phase D — fertility/caregiving. Unlocks caregiver credits (anchor 2's most progressive provision: 52–62% of gains to the bottom lifetime-earnings fifth).
• Phase E — projection + revenue. Trustees-assumption projection and payroll side; unlocks the actuarial-balance Rosetta stone (anchor 2's A-tables) and full-package scores.

Protocol notes for any replication run

1. Baseline convention matters: DYNASIM tables use two current-law baselines — scheduled (full formula) and payable (post-exhaustion cut). Every replication must state which denominator it matches.
2. Replication targets are incidence patterns with documented assumption deltas, not point-forecast matching — per the paper's domains-of-validity doctrine (different Trustees vintages, base data, and behavioral assumptions are named deviations, not failures).
3. Anchors are PDF tables (no machine-readable versions exist); transcriptions above were made from pdftotext output of the fetched PDFs and should be re-verified against the PDFs when a phase registers its gate.

Full per-provision catalog with transcribed table rows available on request (research pass 2026-07-07); source PDFs archived locally.

Start by reading gates.yaml, especially the gate 2 holdout basis, and review the staged files under ~/PolicyEngine/psid-data. The issue defines a multi-phase replication program rather than a bounded change, so a contributor would need to select and scope one phase before implementation. Done would require a registered gate and documented comparison against the named DYNASIM anchor.