Build G: pin-bump validation pass (pe-us 1.752.2 → 1.764.0) before the next release build

[MaxGhenis]

Build G opens with a pin-bump validation pass, per discussion with Max (2026-07-06).

Opener (committed): pin-bump validation pass before any build

policyengine-us moved 1.752.2 → 1.764.0 (28 releases) during the Build F campaign. Before Build G commits to a pin:

1. Resolve the current policyengine-us in a fresh env and run, against the Build F base artifact:
   • the eCPS parity gate (ratchet mode) — expect legitimate state-income-tax deltas from the 1.756.9 uprating-placement fix (33 frozen parameter files across ME/NE/ND/NY/OH/RI now uprate); triage each delta into the reason'd register deliberately,
   • the engine-asserted take-up contract (Seed TANF and EITC take-up from administrative rates; engine-asserted take-up contract (#312) #315) — confirm no model-side seeding appeared for data-seeded flags,
   • critical-fit spot checks (income tax, SS) — quantify model-drift movement before calibration sees it.
2. Record the resolved version; checkpoints invalidate automatically (version is in the cache key by design, Make target_compilation survivable across worker restarts (#299 prereq 1; implements #217) #317).

Candidate program (proposed, not yet committed)

• Accuracy-exemption cluster from the parity register: qualified_tuition_expenses essentially unimputed (641 records, $4.5B) — education credits understated ≥40% #253 (tuition), qualified_passenger_vehicle_loan_interest is never imputed — OBBBA auto-loan deduction is structurally $0 #252 (auto-loan interest), AMT understated ~59%: amt_income carries no preference items (ISO spread, PAB interest, depletion) and the data has no such inputs #254 (AMT preference items), QBI base −31%/−40% vs targets and JCT; add per-component repeal-revenue benchmarks (amount residuals can't see yield gaps) #298 (QBI benchmarks).
• Ingest ASEC 2024 and 2025 into the US support pool (spine is two vintages behind) #296 ASEC 2024+2025 ingest (spine two vintages behind).
• SNAP work-requirement input carriage (SNAP OBBBA work requirements: restore data inputs needed for household exposure analysis #248/Carry SNAP work-program participation inputs or assumptions through Populace US outputs #249/Carry local geography needed for SNAP ABAWD waiver modeling #250, with Derive hours-worked inputs from ASEC reported hours #293/Assign SNAP take-up from reported receipt and the FNS participation rate #294): newly elevated — pe-us 1.758–1.764 shipped the ABAWD/work-requirement engine hooks (exemption inputs, workfare hours, borough-level waivers by county FIPS, HR1 state coverage), which consume person-level inputs populace does not yet carry; HR1 SNAP analyses run on proxies until it does.
• CHIP enrollment targets are concept-mismatched: CMS total_chip_enrollment includes M-CHIP; model materializes separate CHIP only (20 states zero-support) #321 CHIP concept mapping (M-CHIP vs separate CHIP at the ledger level).

Sequencing and scope decisions for the program remain open pending Build F's gate tables (#299).

[MaxGhenis]

Build G opener — pin-bump validation pass (pe-us 1.752.2 → 1.764.0)

Validation-only pass per this issue. No build runs, no register commits, no pushes to main. All compute detached; the battery ran against the reused Build F base artifact base-f-20260705/base_populace_us_2024_puf_support.h5 (sha 18833fb6…, verified before use; 3-year ASEC pool → 2× PUF clone = 337,704 hh). Worktree build-g off origin/main (HEAD e179ba5, which includes #319); runtime scratch in ~/PolicyEngine/_buildg-runtime/.

Resolved versions

• Incumbent pin (Build F / uv.lock): policyengine-us 1.752.2, core 3.26.11.
• uv sync --all-packages --extra us resolves to 1.752.2 — the lock is honored; uv sync alone does not move the pin (that needs uv lock --upgrade). The pin spec is >=1.745.0,<2.
• New-pin candidate: 1.764.0 (this issue's named window ceiling). The spec's live ceiling under <2 is 1.765.3 (current PyPI latest) — 28 releases shipped in the window; battery targets 1.764.0, 1.765.x-only changes listed separately.
• Critical-fit run in two isolated venvs, both pinned to core 3.26.11, so the only moving part is policyengine-us (clean single-variable drift).

---

1. eCPS parity gate (ratchet mode) — PIN-INVARIANT; zero new gaps from the bump

The gate compares input-column nonzero shares of the (frozen) base against the frozen 158-var incumbent-eCPS reference. Because the base H5 is frozen, candidate shares are pin-invariant; the only pin-sensitive input is the engine input-variable scope (841 → 863 vars). None of the 24 new engine inputs are in the frozen reference, so none can create a parity failure.

Confirmed empirically — ran the actual parity_gate at BOTH pins; identical result:

	1.752.2	1.764.0
reference-populated layers	158	158
register exemptions	110	110
unexempted gaps	2	2 (same two)

The 2 unexempted gaps are pre-existing, not pin-induced — both ACA-family, both populated by the native ACA source stage (source_stages.json, #76/#168) which runs before the parity gate in the release, so Build F's release passed while the raw base does not carry them:

layer	ref share	candidate	in register?
takes_up_aca_if_eligible	67.1%	0%	no
selected_marketplace_plan_benchmark_ratio	100%	0%	no

The state-income-tax movement this issue anticipated does not appear in the input-column parity gate — state_income_tax is a computed output, not an engine input. It is quantified in §3 (that is where the uprating fix actually shows up).

Register-entry proposals (PROPOSALS ONLY — lead gates; nothing committed):

• No pin-attributable entries (the bump adds zero gaps).
• Two pre-existing ACA-family exemptions proposed (reason'd, → Migrate ACA health-input gates after PE-US take-up contract changes #76): exempt takes_up_aca_if_eligible and selected_marketplace_plan_benchmark_ratio by name, noting they are populated by the native ACA stage before the gate. Alternative the lead may prefer: measure candidate shares on the post-ACA-stage frame rather than the raw base, in which case no entry is needed. Flagging the stage-ordering, not deciding it.

---

2. Take-up contract assertion (#315) — PASS, all 13 flags @ 1.764.0

Ran the engine-asserted take-up contract tests at the new pin: 14/14 pass (incl. the field-by-field engine-fact check and the "no model_simulated flag is seeded" consistency guard).

Per-flag verdict at 1.764.0 (engine facts vs the checked-in contract, built against 1.752.2):

• 13 flags present — no additions, no removals (no new takes_up_* variable appeared).
• All 13 remain data_seeded (default=True, no formula) — zero migrated to model_simulated. The model-side take-up migration referenced in Seed take-up inputs across programs (parity with policyengine-us-data's stochastic flags) #312/Seed TANF and EITC take-up from administrative rates; engine-asserted take-up contract (#312) #315 still has not landed at 1.764.0.
• Every asserted engine fact (entity, value_type, default, engine_class) matches field-by-field. No model-side seeding appeared for the data-seeded flags; no default-semantics shift.

treatment	flags
seed	takes_up_eitc, takes_up_tanf_if_eligible
out_of_scope	takes_up_snap_if_eligible, takes_up_aca_if_eligible
near_universal	takes_up_medicare_if_eligible
rate_unsourced	medicaid, chip, basic_health_program, ssi, dc_ptc, head_start, early_head_start, housing_assistance

---

3. Critical-fit spot checks — pure model drift on the base, BEFORE calibration

Computed on the base at both pins over a deterministic 40,000-hh subsample (Microsimulation.subsample(n=40000, seed=42, quantize_weights=True)). The full 337,704-hh sim is dominated by pe-us per-household RNG re-seeding (stochastic take-up/imputation), ~7 min per aggregation; the fixed-seed subsample draws the same households at both pins, so the between-pin delta carries zero sampling noise, and weight-quantization keeps aggregates at full-base scale. Percentage drift is scale-stable. (Base is uncalibrated — magnitudes are pre-calibration; the deltas are the signal.)

Federal — negligible, well-behaved drift (all < 0.2%):

target	1.752.2	1.764.0	Δ	Δ%
Federal income tax — amount	$2,000.25B	$2,004.04B	+$3.79B	+0.190%
Federal income tax — returns (>0)	122.20M	122.19M	−0.01M	−0.008%
Taxable Social Security — amount	$466.94B	$467.46B	+$0.51B	+0.110%
Taxable Social Security — returns (>0)	28.069M	28.079M	+0.010M	+0.036%
EITC — amount	$30.822B	$30.808B	−$0.014B	−0.045%
EITC — returns (>0)	13.549M	13.546M	−0.003M	−0.025%
Refundable CTC (ACTC) — amount	$21.614B	$21.619B	+$0.005B	+0.025%
ACTC — claims (>0)	10.431M	10.431M	0.000M	0.000%

The +0.19% federal income tax is consistent with the 1.756.x federal-base changes (non-Sch-D cap-gain-distribution inclusion 1.756.2, OBBBA clean-energy-credit termination 1.756.7, Saver's Credit thresholds 1.756.6) roughly offsetting. Supporting aggregates: AGI +0.17%, taxable income +0.21%, SNAP +0.29% (from the SNAP work-requirement changes), employment income / capital gains / TANF flat.

State income tax by state (the #324-predicted movers) — one material mover:

state	1.752.2	1.764.0	Δ	Δ%	attribution
ME	$2.698B	$2.713B	+$0.015B	+0.55%	1.756.9 uprating unfreeze (12 param files)
NE	$2.826B	$2.827B	+$0.001B	+0.04%	1.756.8 explicit 2026 + 1.756.9
ND	$0.213B	$0.214B	+$0.001B	+0.34%	1.756.9 uprating unfreeze (5 files)
NY	$43.56B	$55.98B	+$12.42B	+28.51%	1.764.1 IT-196 §68 overall-limitation wiring (+ 1.756.8 phase-out unfreeze)
OH	$11.14B	$11.16B	+$0.017B	+0.15%	1.756.9 uprating unfreeze (1 file)
RI	$2.476B	$2.477B	+$0.001B	+0.03%	1.756.8/1.756.9
ALL states	$468.14B	$490.41B	+$22.27B	+4.76%	NY dominates

Headline: NY state income tax +28.5% (+$12.4B). This is not the modest uprating effect — traced to 1.764.1: "Wired New York's orphaned itemized-deduction phase-out parameter into the IT-196 Line 40 pre-TCJA overall limitation (26 U.S.C. 68 via NY Tax Law 615)." Confirmed on-disk: the NY itemized/phase_out/start.yaml parameter changed and the NY itemized-deduction variable sources changed (ny_salt_deduction.py, pre_tcja/*). It reduces NY itemized deductions → raises NY taxable income → raises NY tax. Directionally correct (NY was omitting the §68 overall limitation, under-taxing high-deduction filers) — a legitimate correctness improvement, but large enough that Build G's calibration starts from a materially different NY (and national state-tax) position than Build F did. The other five states show exactly the small uprating-unfreeze moves this issue anticipated.

I confirmed the 1.756.9 uprating fix mechanistically on-disk (the exact "uprating: moved from sibling-of-values: to metadata:" diff): 30 income-tax parameter files changed across ME/NE/ND/NY/OH/RI (ME 12, NE 5, ND 5, NY 1, OH 1, RI 6), the balance of the changelog's "33 files / 107 nodes" being RI/NY contrib reforms.

---

4. New input surface (1.753–1.764) beyond the SNAP/ABAWD wave

Engine input variables 841 → 863 (+24 added, −2 removed). All 24 added default to False/0 and base-F populates none (all consume engine defaults) → degenerate-gate (#286) candidates the moment populace targets the owning program.

• SNAP/ABAWD wave (6, already tracked SNAP OBBBA work requirements: restore data inputs needed for household exposure analysis #248/Carry SNAP work-program participation inputs or assumptions through Populace US outputs #249/Carry local geography needed for SNAP ABAWD waiver modeling #250): is_snap_abawd_discretionary_exempt, is_snap_workfare_participant, weekly_snap_work_program_hours, has_applied_for_unemployment_compensation, is_complying_with_tanf_work_requirements, is_in_substance_use_treatment_program (person-level). This is the SNAP work-requirement input carriage already scoped for Build G.
• State childcare CCAP/CCDP enrollment (13, NEW beyond the SNAP wave): ND/MT/NM/NV enrollment (*_enrolled, spm_unit) + provider-type/quality/hours (person, mostly enums). Releases 1.756.0/1.761.0/1.762.0/1.763.0. Low fiscal risk (opt-in, default off → the program simply does not fire); relevant only if populace targets these state childcare programs — then they are Gate releases on input columns stuck at the engine default #286 candidates.
• Local income tax (4, NEW): in_yonkers, ny_yonkers_nonresident_earnings, in_wilmington, de_wilmington_nonresident_earnings (1.757.0).
• Capital gains (1, NEW): schedule_d_capital_gain_distributions (1.756.0) — the one new input that flows into the federal income-tax base. Base carries 0 at both pins (a definitional-consistency item, not a between-pin mover on this base). If populace later imputes cap-gain distributions, carry it and the pre-existing capital_gain_distributions together.
• Removed (2), no action: small_area_fair_market_rent, safmr_used_for_hcv — moved from stubbed inputs to parameterized/computed (1.754.0); they leave the input scope cleanly.

---

Go / no-go recommendation

GO — pin Build G to 1.764.x, with two documented follow-ups.

The bump is safe and mostly beneficial:

• Parity gate: unchanged — zero new gaps (proven at both pins).
• Take-up contract: PASS, all 13 flags — no model-side seeding appeared, no semantics shifted.
• Federal critical-fit: negligible drift (< 0.2% on every target) — calibration's federal starting point is essentially unchanged.
• State critical-fit: the only material mover is NY (+28.5%), and it is a legitimate correctness fix (the §68 IT-196 overall limitation NY was previously omitting), not a regression. The five other named states move by the small amounts the uprating-placement fix predicts.

Follow-ups the lead should gate, none blocking:

1. Expect NY / national state-tax targets to re-seat. NY state income tax rises ~28% and the national state-income-tax total ~4.8% purely from the model. Build G's calibration and any Build-F-vs-Build-G results comparison must attribute this to 1.764.1, not to data or weighting changes. If a state-income-tax fiscal target is in the feed, confirm it reflects the post-§68 NY (else calibration will fight a correct model).
2. Two pre-existing ACA parity-register gaps (takes_up_aca_if_eligible, selected_marketplace_plan_benchmark_ratio) — add reason'd exemptions (→ Migrate ACA health-input gates after PE-US take-up contract changes #76) or measure parity on the post-ACA-stage frame. Independent of the bump; surfaced by this pass.
3. New-input watch (not now): the 13 state-childcare + 4 local-tax + 1 cap-gains-distribution inputs are default-consuming; add to the Gate releases on input columns stuck at the engine default #286 degenerate-gate watch if/when populace targets those programs. The 6 SNAP/ABAWD inputs are already the scoped Build G SNAP-carriage item.

If the lead prefers to ride the very latest, the same battery should be re-run at 1.765.3 before committing — 1.765.1/1.765.2/1.765.3 add PA/NJ/CO/CA/AZ state-tax fixes not covered here (none touch populace's input surface, but they would move the state critical-fit further).

---

Battery artifacts (all under ~/PolicyEngine/_buildg-runtime/, staging/local only — prod HF and latest.json untouched): probe_pin.py, probe_state_tax.py, out/probe_{1752,1764}.json, out/state_tax_{1752,1764}.json, out/parity_scope_{1752,1764}.json, out/drift_table.md, out/register_proposals.md. Reproduction: base sha 18833fb6…, core 3.26.11 both pins, subsample n=40000 seed=42, period 2024. Full working log in PROGRESS_BUILDG.md on branch build-g.

[MaxGhenis]

Pin committed and validated on the full build — 1.764.6

The Build G certification run (full three-way report on #299) resolved and exercised the pin end-to-end. Recording the outcome against this issue's GO.

Resolved pin: policyengine-us 1.764.6 / core 3.26.11 — uv lock --upgrade-package policyengine-us==1.764.6 then uv sync --all-packages --extra us. 1.764.6 is the latest 1.764.x (the GO's named window ceiling was 1.764.0; 1.764.1–1.764.6 shipped since, all within >=1.745.0,<2), and it includes the 1.764.1 NY IT-196 §68 wiring this issue flagged as the headline mover.

GO follow-up #1 (NY / national state-tax re-seat) — discharged. The opener predicted NY state income tax would rise ~28.5% purely from the model (1.764.1) and warned calibration must attribute it to the model, not fight it. On the full build it calibrates cleanly: NY income-tax liability lands at +0.18% of its SOI target (est $189.2B). The model's post-§68 NY is being hit, not fought — the state-income-tax target reflects the post-§68 NY, exactly as the follow-up required.

GO follow-up #2 (ACA parity register gaps) — not re-encountered on the release path. The two pre-existing gaps (takes_up_aca_if_eligible, selected_marketplace_plan_benchmark_ratio) are populated by the native ACA source stage before the parity gate in the release, so all three artifacts passed the eCPS parity gate; no register entry was needed for the build. (The raw-base-only gap you surfaced stands as a separate note.)

Take-up contract: unchanged behaviour on the build — but note the degenerate-input register did need the refresh flagged in #330: the now-#315-seeded takes_up_eitc / takes_up_tanf_if_eligible exclusions were retired (they carry signal now) and the un-imputed second_home_mortgage_* trio added (→ #38). These were the exact un-merged Build F edits; applied on build-g-run.

Critical-fit at the new pin (on the calibrated build, not just the subsample): federal income tax and SS came in essentially exact on the frozen-57k headline (+0.5% / −0.1%) and dense (+1.2% / +0.1%) — consistent with the opener's <0.2% pre-calibration federal drift. The pin is safe and beneficial, as the GO concluded.

Full gate tables, the cold-L0-2026 head-to-head, and the export-mass identification analysis are on #299. Staging-only; prod HF and latest.json untouched.