[MaxGhenis]

Max presents this deck at the APPAM 2026 roundtable "State Policy Impacts on Family Resources: Tools to Improve Policy Analysis". The roundtable meets on Thursday, November 5, from 10:15 to 11:45am ET in the Wellesley room, and he has 10 to 12 minutes. On November 4, he asked us to publish it at policyengine.org/slides/appam-2026 and keep editing after that, so later edits land in follow-up PRs.

Slides

The deck in slideshows/appam-2026 runs thirteen slides:

1. Cover
2. What PolicyEngine computes
3. One family across 51 jurisdictions
4. Which budget lines differ, with the Texas subsidy and SNAP interaction
5. Benefit cliffs by earnings
6. A real parameter file
7. Simulated eligibility on your own microdata
8. A Massachusetts reform
9. CliffWatch and Build, two of our web apps
10. Where PolicyEngine sits beside the other panelists' tools
11. What we leave out
12. The Axiom Foundation's Georgia SNAP rule, run on the Axiom engine
13. Close

We wrote the talk in SPEAKER-NOTES.md, and in EVIDENCE.md we trace each claim to a computed result or to code at a named commit. We computed every PolicyEngine US figure with 2.90.2 at 469ad12, apart from the 2.2.1 runs, using the scripts in analysis/. For slides 3 to 5 we run one family, and for slide 7 we sample the default Populace US file. We run slides 8 and 11 on the Populace US local-area file, which policyengine.py uses for state regions, and keep their default-file runs in data/*_default.json. For their effective samples, we count each survey household once, merging an ASEC record with its tax-data clone (analysis/effective_samples.py). Slide 12's figures come from the Axiom Rules Engine 0.2.2, and the Census rates from P60-290 Table 18.

The PDF

The moderator asked for slides by Wednesday evening, November 4, to load on a laptop. Max sends the link, and we export a PDF if he needs one to post or if the room's laptop can't reach the site. Export it from the deployment that serves the slides you want, then finish it:

bun run export appam-2026 raw.pdf --url <deployment URL> --size 1440x810 --scale 2
uv run slideshows/appam-2026/analysis/finish_pdf.py raw.pdf Ghenis-PolicyEngine-APPAM-2026.pdf

finish_pdf.py trims each page to its image, links the addresses the slides print, and names the deck and its author in the file. On November 3, we exported the current PDF from Vercel deployment dpl_9X9zS2oSQnTQFbxgNbo27UVnmcxF of e07b7b8, whose slides match 2fb11e9's, and finished it with finish_pdf.py at 2fb11e9. We used headless Chrome 156 and pypdf 6.21.4. The finished file hashes to sha256 dd073b554414349b236e43acb0409a9013109830254a7285f9135d9a70806923, and the raw export to 080e09f9c86c00cdbd9c82ef021e66f9239c8ef532d4768b3b90f60af4a5123a. After 2fb11e9, we changed two slides in the web deck alone. We wrapped the closing slide's QR code and address in anchors without moving either out of the rectangle its PDF link covers. On slide 6, we now cite the 2023 act's sections as main's amount.yaml does, while the PDF's excerpt still reads "Section 22 & Section 45", as in 2.90.2. We then added slide 9, so the current PDF runs twelve pages and lacks it. finish_pdf.py now expects thirteen pages.

Invariants

• Net resources reconcile with household_net_income and spm_unit_net_income to within two cents, across random states and earnings (analysis/test_invariants.py).
• Slide 5's "up to" holds across all 51 jurisdictions over the $15,000 to $100,000 range its caption names. cliff_ranking records each state's largest one-step drop in that range, and the slide lists the top three.
• The sampler gives a household it draws twice fresh ids, so each fixed-sample entry counts once (analysis/test_sample_families.py).
• Each fixed-sample run records the state the model read, and mean employment income holds constant across all 51 runs. State income tax comes to zero in Alaska, Florida and Texas and exceeds zero in Alabama and California.
• No household loses under the Massachusetts reform, and the poverty gap falls by no more than the state spends on it. The child poverty rate falls only when some children's records cross the poverty line, and here it falls (content.test.ts).
• The figures typed into the speaker notes and the ledger match the runs' data, and six that they repeat match at every copy. The ledger's rank for the size of Massachusetts's gap from Census matches the state baseline run (content.test.ts).
• We answer a question on sampling error in the speaker notes and verify three comparisons in content.test.ts. Massachusetts's standard error exceeds twice Census's margin of error there. A normal error of that size would miss Census by more, on average, than our state estimates do, and the median state holds fewer effective households with children than Massachusetts.
• state_file reads blank parent links as 0 and keeps every other id. The model reads 0 as an unknown parent, and any nonzero id that matches no one as an absent parent (analysis/test_state_baseline.py).
• We put state payroll taxes and paid leave contributions on the state line, and the lines add up to net resources (analysis/test_invariants.py).
• Each effective sample stays between one and its record count, and equals the count when weights match. Counting each source household once never raises the sample and keeps it between one and the number of sources. When no two records share a source, it returns the record-level sample, and record order and weight scale leave it unchanged (analysis/test_state_baseline.py).
• effective_samples.py stops unless records that share a source household pair one ASEC record with one tax-data clone whose members have the same ages. It keeps matching ids on the local-area file's two spines apart (analysis/test_effective_samples.py).
• A group-quarters record needs both the record type and the relationship that policyengine.py names (analysis/test_state_baseline.py).
• Standard errors stay nonnegative, scale with their values, vanish across identical households and count a household once however its people split its value. Each SPM unit, tax unit and person takes the source of the one household its members share, in any member order. A unit whose members span two households stops the run (analysis/test_standard_errors.py).
• Every finished PDF page fills 16:9, and each link stays inside its slide's page, over the ink of the address it opens. An export with another page count or size, or a page that already has links, stops the script (analysis/test_finish_pdf.py).
• Slide 4's end labels count DC apart from the states, and slide 7's name both states that tie (content.test.ts).

Differential checks

• On the default file, a Massachusetts cut returns the national run's household gain, gaining share and poverty rates (data/massachusetts_national_check.json).
• The 51-state rates on the local-area file match the separate run in Trace how calibration, sampling noise and rent imputation move state child SPM poverty rates microcosm#1445 to within 0.000001 points in 48 of 51 jurisdictions. That run fills blank parent links with −1 and this one with 0, which moves California 0.56 points, Delaware 0.006 and Missouri 0.00001 (content.test.ts).
• The fixed sample rebuilds to the same rows. Restoring the stored computed variables it leaves out returns the same SNAP and child Medicaid or CHIP shares in Tennessee, Oregon and Hawaii (analysis/audit_sample_columns.py).
• The code slides 7 and 8 print returns the Arkansas and Maryland shares and the Massachusetts cost to every digit (analysis/check_slide_code.py). content.test.ts matches each printed line to that script, apart from the sample file and the state loop.
• Slide 6's values end at the amount slide 8 doubles, and slide 12 prints the 20% rate the Axiom engine applied (content.test.ts).
• A ratio's standard error agrees to within 2% with a jackknife that drops one household at a time (analysis/test_standard_errors.py).
• On a four-person file, each SPM unit's and tax unit's cluster matches the household that policyengine-core reads from its first member. map_to from households to SPM units would instead split household 7's id into 7/3 and 14/3 (analysis/test_standard_errors.py).
• Wisconsin's 2026 subsidy from the schedule run equals the family run's (analysis/wisconsin_schedules.py, content.test.ts).
• The same reform on the same records at policyengine-us 2.2.1 returns $657 million, which the ledger reports beside the deck's $577 million (data/massachusetts_reform_221.json).
• We recompute the median marginal rates from the sweep, and slide 5 prints the sweep's $500 step. The Census rates match the ones the cpid-webinar-2026 deck maps (content.test.ts).
• The precision run recomputes the slide's child rate, fall, poverty gap and cost from its own sums over people, SPM units and tax units. content.test.ts holds it to the reform run's JSON, file, model version and parent-link fill. It clusters each person, SPM unit and tax unit by the source household of the one household its members share, and stops if a unit's members span two households. We treat the calibrated weights as fixed, so the standard errors leave out what calibration adds. They also leave out errors that every record shares, such as a data-wide error in rent, and those that every clone shares through the model that imputes its tax detail (analysis/massachusetts_precision.py).
• Record-level Kish values in data/effective_samples.json reproduce the simulation runs' values to within one part in a million on both files. Massachusetts's source-level values match the reform run (content.test.ts).
• On each page, the PDF holds one link for each anchor its slide renders in the web deck, and the slide prints every address its PDF page links to (content.test.ts).

Slide 2

Slide 2's reform line reads "Who would gain or lose if lawmakers adopted a proposed reform." The session blurb's line reads "How proposed reforms would change poverty and household incomes in each state," and Max can switch the slide back to it with one edit in WhatItDoesSlide.tsx.

Slide 9

Slide 9 shows two of our web apps. In CliffWatch, we entered a single parent, 33, with children aged 6 and 10 in Cambridge, Massachusetts, and kept its default programs. In Build, we set the Massachusetts child and family tax credit amount from slides 6 and 8 to $880 from 2026. We captured both from the live sites on November 4. For Build, we set the browser's time zone to UTC: in New York time, Build labels each gridline in its chart of historic values with the year before.

We keep CliffWatch's figures off the slide. CliffWatch runs PolicyEngine US 1.755.5 and counts Medicaid and CHIP in net income. We computed the deck's figures with 2.90.2 and leave health coverage out of net resources. In the speaker notes, we cite one CliffWatch result: this family loses TAFDC, Massachusetts's cash assistance, when its earnings pass $18,000.

Review

Reviewers in rounds 1 and 2 requested changes, and we recorded what we changed in RESOLUTION-r2.md. In round 3, five reviewers read the deck at 0129c2a, one of them the PDF alone. Skeptics then tried to refute each finding, and a critic looked for gaps. The skeptics confirmed 89 findings, four of them must-fix, and refuted two. Three findings needed a microsimulation, which we ran on November 3: we now cluster the standard errors by source household (A3), and the column audit lists 181 undefined names (A11). We also reran the reform with blank parent links read as 0, recorded that fill, and found slide 8's figures unchanged (A7).

On November 5, an Opus 5.5 reviewer approved slide 9's commit (8d86641..236e809). It replayed CliffWatch's household on the live API, compared the screenshots with the capture run's stills, checked every slide reference and ran vitest and pytest. Its four smaller findings go in a follow-up PR.

With the Codex lanes closed, an Opus 5.5 reviewer then read the deck at d07f110 and requested changes, with one must-fix, five should-fix and six optional findings. Max decides one of them: whether slide 8 keeps its $169 million poverty gap card or moves that figure into the footer. We fixed or answered the rest, and recorded what we did with each round-3 and final finding in RESOLUTION-r3.md. Both resolution files sit in ~/reviews/appam-2026/ on Max's machine.

We now count each survey household once in the effective samples. Massachusetts's falls from 395 to 301 households, or from 173 to 132 among those with children. Among households with children, the median state's falls from 125 to 108 on the local-area file and from 75 to 63 on the default file. We also fixed a bug that had mislabeled the SPM-unit and tax-unit clusters. With the source-household clusters and that fix, the standard errors rise to $64 million on the cost, $42 million on the gap and 4.6 points on the 2026 child rate. They had read $54 million, $34 million and 3.8 points.

Two verification reviewers then read the deck at 2fb11e9. The statistics reviewer approved it and asked for a test that holds the cluster fix, so we added one to analysis/test_standard_errors.py. When we reran analysis/massachusetts_precision.py, it wrote the same data/massachusetts_precision.json. The PDF reviewer could open the page images but not the text files, so it checked how the pages look and found nothing to fix there. We extracted each PDF link and cropped the page image under it for a second PDF reviewer.

On November 4, that reviewer approved the deck at 8f07254. It compared all 12 page images with the slide source and found no differences, checked all nine link boxes, and asked for no new export. We fixed its findings on the closing slide's anchors, the link test and two ledger sentences in 8d86641, and corrected two sentences in this description. It had no network, so we checked the outside sources ourselves: the TAXSIM cutoff, the dashboard's reform options, the 2035 horizon, the RuleSpec excerpt and the Department of Labor's wage table. Each matches the deck and the ledger. Another reviewer then checked those fixes and our source checks. It asked for no change to the PDF and found no fault in the deck, and we corrected the records it flagged.

Impact

The PR adds one registry entry in lib/slideshows.ts and the QR image and two screenshots in public/images/appam-2026/, and everything else sits in the deck folder. GitNexus impact failed on a storage version mismatch, and its change detection reads the main checkout rather than this worktree. So we checked the scope by hand, and git diff --stat origin/main shows no other shared code.

axiom: n/a: slides and analysis, no change to a country model

🤖 Generated with Claude Code

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
policyengine-slides	Ready	Preview	Nov 5, 2026 2:50am UTC

[MaxGhenis]

Max asked on November 4 to publish this deck at policyengine.org/slides/appam-2026 so he can send the moderator the link, and to keep editing after that. Merging at 236e809.

• gh pr checks exits 0 on this head, and GitHub reads the PR as mergeable with no changes requested.
• An Opus 5.5 reviewer approved slide 9's commit (8d86641..236e809) with nothing to fix before publishing. GPT-6.1 Sol approved the deck at 8f07254, and 8d86641 fixed that review's smaller findings.
• The slide 9 reviewer's four smaller findings and the speaker-note fixes land in a follow-up PR, since this merge also opens the deck's address to edits from main.