Conversation
estimates/group_breakdown.py tabulates any blind test's person rows by SSA's MINT8 row groups (sex, race and ethnicity, country of birth, age, marital status, education, poverty status, household income quintile, benefit type) and by data-driven alternate schemes. It reports the tests' own statistics (ratio of scenario means over draws, poverty change, shares) beside MINT's (percent with a decrease or increase at the 1 percent threshold, 10th/50th/90th percentiles of individual percent change, poverty counts). Group floors restrict the full-sample family split; suppression flags mark cells under SSA's 100-case rule without dropping them. The MINT8 labels come from captured SSA sources pinned by SHA-256. Invariants tested: groups plus unclassified partition Total; ratio statistics are invariant to weight scaling; Total rows agree with the existing COLA, uniform-cut and Track M tabulators; percentiles are bounded; decrease + unaffected + increase = 100. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
estimates/lifetime_measures.py computes, from cohort outputs only: current-law AIME at age 62 (MINT8's initial AIME), the present value at 62 of OASDI payroll taxes, own and shared between spouses for married years (MINT8's lifetime payroll tax measures), the Butrica-Uccello average of wage-indexed earnings at ages 22-62, and weighted quintiles with MINT's labels, cut per birth cohort or over the whole population. OASDI tax rates and trust fund effective interest rates are captured from SSA with SHA-256 pins. Unrecorded conventions are exposed as named builder defaults for the registration to fix. Invariants tested: quintile weight bounds and monotonicity; invariance to weight scaling and row order; zero earnings give zero tax; tax is linear below the cap and flat above it; sharing conserves a couple's total; AIME agrees with the ss oracle. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
data/group_attributes_psid.py reads, with exact label checks and documented code domains, the family-file race and Hispanic-origin reports of reference persons and spouses, the individual file's completed education by wave, and country of birth where the PSID asks it. cohorts/group_attributes.py turns them into a person-keyed side frame for the 2009 and 2011 projection cohorts, the age-67 observations and the Track M 2023 universe, mapped to MINT8's categories and to the Butrica-Uccello report rows. Conflicts across waves resolve deterministically; people never asked (other family-unit members) get an explicit, counted unknown. The side frame has its own file audit and SHA-256 seals, so the existing cohort frames and their pinned digests do not change. Invariants tested: every requested person keeps exactly one row; labels, domains and source pins are enforced; resolution is deterministic; mutation is detected. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…-r2' and 'nasi/g3-group-tabulation' into nasi/g-base
group_breakdowns/cola.py and fra68.py re-execute each frozen registered specification (projection, benefits, union rows) by composing the existing modules read-only, and refuse before loading any group attribute unless every committed cell and per-draw diagnostic matches runs/replication_urban2010_cola_v1.json or runs/replication_urban2010_fra68_v1.json. They then join G1's race, education and nativity side frame and G2's lifetime measures by person_id and tabulate with G3 for both the tests' own population and MINT's beneficiaries aged 60 or older in 2030. Poverty status and household income quintile are not computable in the projection (no income) and are marked so. scripts/run_projection_groups_registered.py refuses without an issue #42 pointer, a clean HEAD equal to the registered commit and an absent output; scripts/make_nasi_repro_venv.sh builds the environment the committed artifacts recorded. Invented dry runs only; the large invented result files are kept outside the repository and the RESULTS.md summaries are committed. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
group_breakdowns/uniform_cut.py re-executes the frozen Track U specification for every registered row (build_age67_cohort -> income_rows -> adjusted_incomes -> tabulation_rows -> tabulate_uniform_cut with the design frame), checks the inputs digest and every committed cell of runs/replication_boomers2004_uniform_cut_v1 with zero tolerance, and refuses before loading any attribute on a mismatch. It then adds MINT groups (race and ethnicity, education, lifetime earnings quintiles own and shared, poverty status, household income quintile) and the report's own rows, with the adjusted-poverty statistic, design standard errors and floors, plus MINT's poverty-table statistics. Every observation must be born 1936-45; the 1946-55 cohort (the blind U2 target) never enters. scripts/run_track_u_groups_registered.py refuses without an issue #42 pointer, a clean HEAD equal to the registered commit and an absent output. Invented data only; the large invented result is kept outside the repository. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
group_breakdowns/min_benefit.py re-executes Registration 17's registered computation (cohort, careers, evaluate and the unchanged tabulate_track_m for every registered row and the d430 sensitivity) and refuses before loading any attribute unless every committed cell, n column, floor, design SE and cohort structure of runs/replication_urban2006_minimum_benefit_v1.json matches. It then adds MINT groups (marital status in MINT order with unclassified counted, MINT age bands from 2022 - birth year, sex, race and ethnicity, education, initial AIME and lifetime payroll tax quintiles, benefit type) to the share receiving the minimum under each option. tests/group_breakdowns/test_min_benefit_reproduction_psid.py is a host-only check that the committed (already published) cells reproduce exactly; it computes no group cell. Invented data otherwise. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…ily from g4a; reconciled in the next commit) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…rarily from g4a; reconciled in the next commit) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
group_breakdowns/common.py is now one shared module for all four adapters (cola, fra68, uniform_cut, min_benefit): one zero-tolerance comparator, one registration preflight (issue #42 pointer, HEAD equal to the registered commit, clean tree including untracked files, a new runs/ destination), one paired exclusive write with rollback, and one environment builder. Where the three builders' helpers differed, the stricter behavior is kept. Every adapter still refuses before any group attribute loads unless its committed artifact reproduces. The three captures of SSA's MINT8 Table User Guide collapse to one canonical capture (certified 2026-04-01) and one label file; tests show the snapshots' main content and labels are identical. The Butrica-Uccello report scheme now follows the cleared exercise-2 definitions extract: race rows, own and shared lifetime earnings (wage-indexed earnings at ages 22-62 over a 41-year divisor, shared in married years). The extract leaves the education and labor-force- experience definitions open, so those rows stay unavailable, and the unrecorded quintile conventions are named choices for the registration. Adds the ten new src modules to the birth-evidence reducer's post- review exclusions and the new tests to the tier manifest. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…6, integration_psid 1,346) Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
|
The latest updates on your projects. Learn more about Vercel for GitHub.
|
…ent years under ALL_AGES Independent review of cd40245 (REQUEST_CHANGES): - Major: the projection and minimum-benefit entry scripts wrote the caller's --output spelling while preflight checked its absolute repository path, so a relative --output from another directory wrote elsewhere and left the one-shot guard's file absent. Both now write state['output_path'] from preflight. Regression tests run preflight from another directory with a relative path and confirm a second preflight refuses. - Minor: under the ALL_AGES divisor, shared report earnings skipped married years absent from the person's own career, dropping the spouse's half (the reviewer's case returned 487.80 instead of 500). Those years now share; the reviewer's case is a test. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
…eakdowns PR Keeps both sides' reducer exclusions. Tier manifest from a full collection, reconciled with master plus this PR's delta: unit 6,401, artifact 3,904, integration_psid 1,346. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
CI on 29e6eea failed two pin tests the targeted runs did not cover: - tests/estimates/test_coordinator.py pins every module under estimates/; group_breakdown.py and lifetime_measures.py are post-compute re-tabulations outside the registered first-estimates surface, declared as their own surface beside Track A's and Track U's tabulators (coordinator._ESTIMATOR_SURFACE_SOURCES is unchanged). - tests/ss/test_statutory_aime.py pins every module that names LEGACY_FIXED_35; lifetime_measures.py uses it for the exercise 1 and 3 AIME conventions, so the projection tests' quintiles use the AIME their benefits use. Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
This branch was successfully deployed
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Builds the machinery for the first NASI follow-up from 2026-10-01 (Kathleen Romig, CBPP): report each blind test's results the way SSA publishes MINT results, by group and not only in total. This PR adds code and tests only. It computes no real-data group result. Each real-data run needs its own registration on issue #42 first.
MINT's categories, matched to the page
SSA's MINT8 Table User Guide ("Definitions—Table Rows and Columns", certified 2026-04-01) defines these row groups for its annual beneficiary tables.
data/external/mint8_row_categories.jsonholds the labels verbatim, taken from the row labels of one payroll-tax option table with no data cell read:MINT8's annual tables have no lifetime earnings quintile. Its birth-cohort tables carry three lifetime measures instead: initial AIME quintile, lifetime payroll tax quintile, and the same shared between spouses for married years. This PR implements those three as the lifetime-earnings dimension.
MINT reports the share of people with a benefit decrease or increase of 1 percent or more, and the 10th, 50th and 90th percentiles of each person's percent change. It reports no means. The tabulator reports those statistics beside each test's own statistic.
What is added
data/group_attributes_psid.py,cohorts/group_attributes.py). Label-verified PSID readers for race and Hispanic origin (family files, 1985–2023), completed education (individual file) and country of birth (from 2013), built into a person-keyed side frame.estimates/lifetime_measures.py). AIME at 62, the present value at 62 of OASDI payroll taxes (own and shared), the Butrica–Uccello average of wage-indexed earnings at ages 22–62, and weighted quintiles with MINT's labels.estimates/group_breakdown.py). Category schemes as data, covering MINT8 and the Butrica–Uccello report rows.group_breakdowns/). One per test:cola,fra68,uniform_cut,min_benefit.runs/artifact reproduces exactly.scripts/make_nasi_repro_venv.shbuilds the Python environment each committed artifact recorded.What is not computed
Protected files
This PR edits none of the following:
engine/loop.pyorengine/steps.pygates.yamlruns/*.jsondata/external/track_u2/u1_identity.jsoncola_track_a/,fra68_track/andss/, orestimates/cola_age_profile.pyInvariants (property-based and example tests)
ssoracle.Tests (targeted, on the host)
tests/group_breakdowns,tests/estimates/test_group_breakdown*.py,tests/estimates/test_lifetime_measure*.py,tests/cohorts/test_group_*.pyandtests/data/test_group_attributes_psid.pypass.tests/estimates/test_birth_evidence_artifact.py,tests/track_u2/test_u2_isolation.pyand the fourtests/test_replication_*.pypin tests pass.axiom: n/a (tabulation and cohort-attribute code; no policy rule changes)
🤖 Generated with Claude Code