[MaxGhenis]

Draft until Max's call, due by 14:45Z on November 5. Main stays untouched until then, and this PR merges only with his OK and an independent APPROVE at the head that merges. If he keeps today's deck, we close this PR as draft work.

This PR carries #99 (4:3 and short-window fit) and #100 (posted child care limits for Idaho and Ohio's voucher), so merging it lands both. It also applies the posted limits Sol's review of #100 asked for: Missouri and Alabama at $3,331 a month, and Indiana at $3,331 before its April limit.

What changes

The talk opens on one Massachusetts family and one question, and widens from there. Fifteen slides in about 12 minutes:

1. Cover
2. One family, one question: how much SNAP in a year?
3. We surveyed AI models, with a build that reveals PolicyEngine's answer
4. The family's other lines, and PolicyBench's rates across 100 households
5. The waterfall from earnings to net resources
6. Across states
7. Across earnings
8. The Massachusetts child and family credit at $880 for this family
9. The same change for the state
10. One model, several ways in
11. CliffWatch and Build recording
12. Coverage: policies, and how we build the data
13. What we're working on
14. Axiom
15. Thank you and QR code

It drops the comparison table and "What we leave out", and it folds the old slides 2 and 4 into spoken lines. Questions keeps the answers those slides carried.

Slide 3's data. Slide 3 reads the registered second round: five calls to each of ten models, tools off, default effort, on analysis/model_ask/prompt-v2.txt. All 50 answered on the first attempt. The slide shows the seven newest versions (Max's rule). Of their 35 answers, 8 came within 10% of $3,663 and none within $1, and the notes give 15 of all 50 and the older versions' 7 of 15. model_answers.py refuses a comparison file whose status isn't final, and no count sits in slide text: each comes from the file.

Massachusetts. Slide 2 asks the room to assume approval for all twelve months, and slide 3's reveal prints SNAP with the subsidy ($3,663) and without it ($8,739). The notes state the assumption at both, and say Massachusetts keeps a waiting list and scheduled new vouchers from October, so a new applicant may not get a full year. The department's counts sit only in Questions.

Slide 6. At $40,000, the family ends up with $26,562 in Indiana, $27,711 in Missouri, $28,166 in Idaho and $32,224 in Alabama. Three of those rows turn on $2 a month, and Questions gives the rows at the limits the model reads: $31,487, $41,931 and $40,722. Slide 7's largest drops become Idaho −$17,808, Maryland −$15,343 and Missouri −$15,124.

Invariants, stated and tested

• Waterfall identity: earnings, taxes, credits, benefits and the child care bill chain to net resources within two cents (content.test.ts, test_waterfall_adds_up_for_this_family, test_waterfall_splits_the_net_tax_lines).
• Doubling the credit moves only the state refundable credit line and net resources (test_doubling_the_credit_moves_only_the_state_credit_line, content.test.ts).
• Monthly SNAP sums to the annual amount with and without the subsidy, and the slide prints that arithmetic (9 × $297 + 3 × $330 = $3,663; 9 × $720 + 3 × $753 = $8,739).
• Monthly program amounts, capped at the family's annual care bill, reconcile with the deck's subsidy line in all 51 jurisdictions.
• Posted limits: each lowers its row against the unmodified model (test_each_posted_limit_lowers_this_familys_subsidy_and_net_resources), and a family under the limit keeps its subsidy (test_a_family_under_the_posted_limit_keeps_its_subsidy).
• Enrolled families: the subsidy never falls when a family already sits in the program, at every grid point in eight states, except Ohio's labelled voucher band from $39,000 to $53,000, where the test asserts the opposite order. A Python comparison covers Ohio's total subsidy on both sides of the band.
• Slide 3's counts come from the calls: content.test.ts recounts the headline and every row of slide 4's table from comparison-v2.json's 50 calls.
• Slide 3's chart draws every answer the seven gave, and the test compares the sorted lists.
• Attempts: the file records no retries, no malformed answers, no unanswered calls and no tool use.
• Slide 7's chart gives each highlighted state a defined stroke of its own.
• Coverage: the local-area file's households and targets add from their parts, and the slide names which file each figure describes.
• Links: finish_pdf.py's table lists each anchor on each slide, measured on this PR's deployment.

Calls made

• Slide 4 drops the score line. The bars read the 2026-10-27 snapshot and the scores read the October 23 copy, so the scores moved to Questions.
• Slide 13 keeps five items, each an open PR, and leaves out microcosm #1248, which reads as a list of gaps.
• Indiana takes $3,331 before April. We found no chart Indiana posted before April 5. Our model pairs 150% with the 2026 guideline, 150% of the 2025 guideline gives $3,331, and the PN3 Roadmap prints 140% for 2025. Chalkbeat reports a 15-month freeze to May. Questions gives the row at the model's limit.
• Alabama's five months count a new applicant in every month. Questions says a family Alabama approves in May keeps the subsidy in June.
• Georgia keeps its CAPS subsidy, and Questions says Georgia denies new applicants outside its priority groups and gives the row without it.
• Fit the APPAM deck to 4:3 screens and to short and wide windows #99's short-window rule stays as it is. Its review asked to restore (max-height: 500px). On production main, 1024x576 clips slides 1, 6, 7, 8, 9, 10 and 12 (least spare −64px) and 1000x520 clips nine slides (−110px), and viewers can't reach that content. With Fit the APPAM deck to 4:3 screens and to short and wide windows #99, they can scroll to it. Neither size is a standard projector mode.

Evidence

See the review folder, ~/reviews/appam-2026/rebuild/deck-review/, for saved test output, the snapshot and the diff at each reviewed head. CI doesn't run the Python suite, so the folder also records its runs on the machine.

🤖 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 2:36pm UTC

[MaxGhenis]

Audit: the coordinating session merged this on Max's go at 7c3e6f8. Reviews: review-r1-opus.md (Opus 5.5, APPROVE at 05812f0), review-head-roadmap-opus.md (APPROVE) and review-r1-opus-deck.md (Opus 5.5, REQUEST CHANGES at 05812f0, which arrived after the merge decision). #102 fixes that review's findings. The Sol review 20261105-091841-rebuild-r1 didn't run, because the Codex lanes were closed. A Python subset stopped after 33 passes, so the full suite still needs a run after the talk.