Document every result's baseline and make the projection baseline selectable (TR2008, TR2026, CBO 2026)

[MaxGhenis]

Answers the third NASI follow-up from 2026-10-01 (Jon Schwabish): state the baseline behind every result, and make it selectable between the latest Trustees intermediate assumptions and CBO's long-term projections, with fertility by single year of age rather than one total rate.

What changes

Documentation (docs/baselines.md, new page in the docs book).

• One table gives the baseline behind each published result: exercises 1 and 3, exercise 2, exercise 4, Registration 18, and gates m6, 2, 2b, 2c and w1.
• Per-assumption sections cover COLA, AWI, mortality, claiming, DI, fertility, earnings and immigration, with code citations.
• A "Choosing a baseline" section and a "Fertility by age" section describe the new package.
• "Known discrepancies" records the A1 §15 AWI wording and the claiming table vintages.

One paper sentence corrected. paper.qmd said the Trustees Report publishes no projected death probabilities by single year of age and sex. The 2008 Report prints none, but OACT publishes them with every Report since 2014. The sentence now says so and cites the 2026 file. No number in the paper changes.

Inputs (data/external/tr2026/, data/external/cbo2026/; accessors in data/tr2026.py and data/cbo2026.py).

• 2026 Trustees intermediate assumptions:
  • V.C1 COLAs for 2025–2035, then V.B1 CPI-W growth (tagged derived);
  • the VI.G1 AWI;
  • the V.A1 total fertility rate and age-sex-adjusted death rates;
  • OACT's probabilities of death by single age and sex, 1900–2100.
• CBO 2026:
  • single-year fertility rates for ages 14–49;
  • mortality by age and sex. Read as probabilities of death, these reproduce CBO's published life expectancies within 0.0011 years; read as central rates, they do not;
  • CPI-U and real earnings growth;
  • covered earnings per worker.
• The 2026 Statistical Supplement's claiming table, with entitlement years through 2025.
• Every file is SHA-256 pinned, with provenance and a transcription check. Pinned source folders are marked binary in .gitattributes, because git's line-ending normalization would otherwise break the pins on checkout.
• No cost, income or balance table was captured.

Selectable baselines (src/populace_dynamics/baselines/). get_baseline(name) returns one of three:

• tr2008_intermediate is the default. It delegates to the constructors the registered tests use.
• tr2026_intermediate uses the 2026 Trustees assumptions. The Trustees publish only a total fertility rate, so single-age rates scale the NCHS 2024 age shape to each year's total, tagged derived.
• cbo2026_long_term uses CBO's published single-age rates.

How the new baselines behave:

• Both splice realized COLAs through 2025 and the realized AWI through 2024.
• DI rates stay 2008-fit under every baseline, a named gap: neither source publishes DI rates by age.
• build_track_a_inputs refuses a real-data run under a non-legacy baseline, because such a run would need its own Candidate 2 design: latent-permanent conditioned chained QRF (research memo) #42 registration.
• baselines/fertility.py adds an opt-in single-age birth step for open populations. It is not gated, it is report-only, and it does not touch the gated fertility law.
• scripts/describe_baseline.py prints any baseline's input paths.

No published number changes. Every registered result keeps tr2008_intermediate. This PR edits no file in runs/, gates.yaml, engine/loop.py, engine/steps.py, or the v1 modules under cola_track_a/, fra68_track/, ss/ and estimates/cola_age_profile.py.

Invariants (property-based and example tests)

• tr2008_intermediate equals the existing constructors exactly. Differential tests cover every COLA (2008–2030), AWI (1975–2085), mortality array (2009–2030), claim pmf and DI rate the registered tests read.
• Every baseline:
  • q(x) lies in [0, 1], and q at age 119 is 1 where the source closes the table;
  • age-specific fertility is non-negative and sums to that year's TFR;
  • the AWI is positive;
  • COLA paths are continuous at splices;
  • claim pmfs sum to 1.
• Life expectancy recomputed from q(x) matches TR V.A4 and CBO sheet 1 within stated tolerances.
• The band interpolation preserves each five-year band's sum and stays non-negative.
• In the birth step, expected births equal the sum of rates over exposed women; zero rates give zero births.

Tests (targeted, on the host)

• tests/baselines/ (114), tests/data/test_tr2026_artifact.py, tests/data/test_cbo2026.py, tests/test_baseline_life_table.py and tests/test_ssa_claim_ages_2026.py: all pass, with 64 unit and 175 artifact tests added to the tier manifest.
• tests/estimates/test_birth_evidence_artifact.py, tests/test_paper_text.py and tests/track_u2/test_u2_isolation.py pass.
• quarto render docs/baselines.md passes.

axiom: n/a (projection-input data, docs and an opt-in baseline layer; no policy rule changes)

🤖 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 4, 2026 6:30pm UTC

[MaxGhenis]

Merge record: merged at reviewed head 278abd4 (merge commit ff3fe30).

• Reviews (independent Subfleet agents):
  • d9bfc0f: Opus, REQUEST_CHANGES, docs only, no blocker.
  • ac4857f: GPT-6.1 Sol executing delta review, APPROVE. All 8 earlier findings resolved; 264 targeted tests pass; tier counts match a full collection; 18 ASFR and 12 q(x) cells match the CBO source exactly.
  • 278abd4: Opus confirmation review, APPROVE. Both minor findings fixed (relative-import resolution, mortality-year example); the merge with master kept both reducer lists intact; manifest equals a full collection (unit 6,029, artifact 3,536).
• CI: every check passed on 278abd4.
• Scope: no published number changes. Every registered result keeps tr2008_intermediate, and no runs/, gates.yaml, engine/loop.py, engine/steps.py or v1 track module was edited.