[MaxGhenis]

What this changes

CLAUDE.md, the README and two docs pages named microimpute, microcalibrate, L0 and the Enhanced CPS as this project's tools. PolicyEngine archived microimpute on 2026-11-08 (PolicyEngine/microimpute#224), and Microcosm's microcosm-fit and microcosm-calibrate succeed those packages, as docs/infrastructure.md and docs/methodology.md already say. This PR brings the four files in line and corrects how docs/infrastructure.md describes microcosm-calibrate's solver.

• CLAUDE.md, core technical concepts. The model now extends Microcosm instead of the Enhanced CPS. Step 1 names microcosm-fit (RegimeGatedQRF) and notes that the engine imports it as populace.fit; step 3 names microcosm-calibrate. A new line tells agents to build new imputation on microcosm-fit and new calibration on microcosm-calibrate, says the microimpute, microcalibrate and L0 packages belong to PolicyEngine's previous data stack, and points to docs/infrastructure.md. The "Key insight" line about the Enhanced CPS matching JCT and TPC goes.
• CLAUDE.md, related repositories. One microcosm row replaces the microimpute, microcalibrate, policyengine-us-data and L0 rows.
• README.md, related projects. A Microcosm link replaces the PolicyEngine-US-Data, Populace, microimpute and microcalibrate links.
• docs/_static/methodology-interactive.html. The pipeline nodes and detail panels name microcosm-fit and microcosm-calibrate, and their code samples now call those packages' APIs: RegimeGatedQRF(seed=0).fit(...) then .predict(...), and calibrate(base_population, ssa_targets) then .frame. The CPS panel describes Microcosm's U.S. build. The page drops its Enhanced CPS claims and the "7,000+ ECPS calibration targets" card, and its key insight opens with the tools PolicyEngine already has.
• docs/plans/employer-firm-plan.html. The diagram node, the data table, the workstream text and the target/gate protocol rule name microcosm-fit and microcosm-calibrate. Two Enhanced CPS references now describe the recipe and PolicyEngine's microdata, and the CPS row drops a third, "(the ECPS host)".
• docs/infrastructure.md. The microcosm-calibrate entry, which the new CLAUDE.md line points to, said "accelerated projected gradient". It now says Adam on log-weights by default, with opt-in L0 sparse selection.

Evidence from code

Claim	Where
The engine imports microcosm-fit's QRF as populace.fit	this repo's master d31f22c: src/populace_dynamics/engine/forward_earnings.py:981 and src/populace_dynamics/engine/refit.py:780 (from populace.fit.qrf import RegimeGatedQRF); PolicyEngine/populace redirects to PolicyEngine/microcosm
RegimeGatedQRF(...).fit(df, predictors, targets, weights=...), then .predict(df)	Microcosm main 3db65154b, packages/microcosm-fit/src/microcosm/fit/qrf.py:1755 (fit; a DataFrame needs explicit weights, :1769-1772), :2349 (predict)
calibrate(frame, targets) returns a CalibrationResult carrying .frame	packages/microcosm-calibrate/src/microcosm/calibrate/solve.py:2118-2151 (signature), :283-336 (CalibrationResult, with frame: Frame at :336)
microcosm-calibrate runs Adam on log-weights by default	solve.py:2124 (method: str = "adam"), :2164-2175 ("apg" survives only as a deprecated alias for "adam")
microcosm-calibrate applies L0 selection only when the caller sets l0_lambda or target_records, and both default off	solve.py:2131-2132 (target_records defaults to None and l0_lambda to 0.0), :2595 (the target_records branch searches l0_lambda)
microcosm-fit runs in Microcosm's U.S. and UK builds	packages/microcosm-build/src/microcosm/build/us_runtime/puf_qrf_chain.py:44,219,363 (QRF from microcosm.fit); uk_runtime/was_wealth.py, lcfs_consumption.py, etb_vat.py and etb_services.py call RegimeGatedQRF, and spi_income.py and salary_sacrifice.py load microcosm.fit.QRF
Microcosm's U.S. build calibrates with microcosm-calibrate	us_runtime/exact_k_ladder.py:19-29,190,217 (calibrate on the release ladder)
Microcosm ships an MIT license	Microcosm LICENSE
The page's nodes still open their panels	every showDetail('…') key matches a data-detail attribute and a detail panel, and node --check passes on the page's script

Not in this PR

• The README's ten remaining populace prose lines (30 to 181). Package rename populace_dynamics → microcosm_dynamics requires a design-binding repin ceremony #373 exempts README prose from the package repin, so a prose pass can rename them.
• The plan page's other Populace names, since the page records a plan from 2026-07-14.
• docs/adr/0003-employer-firm-extension.md:200, which names microcalibrate in the rule this PR rewrites on the plan page. ADR 0003 changes only by joint PR between workstreams A and B.
• docs/infrastructure.md:311-336, which shows a policyengine-us-data/ tree as this project's code layout.
• paper/paper.qmd:454, which cites @ghenis2024 beside @policyengine2026 for PolicyEngine's open cross-sectional microdata. The paper cites @microcosm2026 for Microcosm's frame at line 365.
• reviews/ feedback records, which keep their wording.
• The ghenis2024 citation key at CLAUDE.md:46, which cites the 2024 Enhanced CPS launch post.

🤖 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	Nov 8, 2026 8:19pm UTC