[MaxGhenis]

This PR fixes slide 12's live tracker and slide 15's Georgia caption, adds the record behind Build's figures, and updates the PDF link table. Production runs 7954cfe, and these fixes waited for Max's talk to end.

What changes

Slide 12's live tracker (slides/CoverageEmbed.tsx)

• The deck swaps the still for the tracker only after a poll, run while the slide shows, finds three things: the snapshot line with its commit (Coverage reflects policyengine-us 2.72.3 (main @ 4c65c82), no font still loading, and the four count cards (Total programs, Complete, Partial, In progress) at full opacity and size, ancestors included.
• At 7954cfe the deck checked only the line and the fonts. A WebKit run on production (rc12-dwell-webkit.log) showed the frame 634 ms after the slide arrived, while the tracker's cards read opacity 0.38, 0, 0 and 0 at scale 0.9, so the frame replaced the still while three cards were invisible.
• A hidden frame's cards read as at rest before the tracker fades them in (same log, before arrival), so the deck polls only while the slide shows.
• A reload inside the frame (pagehide) brings the still back until the check passes again, because the tracker prints its fallback counts first.
• On a branch preview, the frame's origin differs from the deck's, the deck can't read its document, and the still stays for good. At 7954cfe a preview swapped in the frame at the first poll.
• The year menu no longer keeps the arrow keys. Every key pressed in the frame reaches the deck unless its target is a text field (input, textarea or contentEditable).

Slide 15's caption (slides/AxiomSlide.tsx) now reads: "Georgia allows the deduction for a household with earned income, and applies the rate to its total countable gross earned income." The old caption required the household to report on time. In section 3650, Georgia denies the deduction only for income reported late, when it computes an overpayment, and 7 U.S.C. 2014(e)(2)(C) and 7 CFR 273.18(c)(1)(ii)(B) set the same limit; the regulation keeps the deduction in an agency-error claim. Axiom's 3612.yaml zeroes the whole deduction and doesn't encode 3650 yet, so the caption states the rate and its base, and the notes' Questions cover the late-report case. The body now opens with "For example," since it features one rule.

Build's figures (data/build_report_4940.json, analysis/massachusetts_reform_national.py, data/massachusetts_reform_221_national.json, EVIDENCE.md). Report 4940's record, saved from api.policyengine.org/us/report/4940, names and hashes the national file, and holds $657,359,316 with child poverty falling from 22.23% to 21.72%. On the local-area file at 2.2.1, we estimate $657,356,800 and a fall from 22.25% to 21.73%. On the national file, we estimate $628.7 million, and child poverty stays at 21.4%. Neither run reproduces the report's 45.7% of people gaining; on the national file, we estimate 46.7% gain more than $1 and 44.5% gain more than 0.1% of household income.

Notes (SPEAKER-NOTES.md): Build's paragraph, three Questions (versions, Build's 45.7% against slide 9's 29.4%, and Georgia's late report), and slide 12's stage direction for the search box and the year menu.

PDF links (analysis/finish_pdf.py): nine rectangles on slides 11, 12, 14 and 15, from the anchors measured on the 31b59c3 preview at 1440×810 (rc14-links.json).

Invariants

1. The frame shows (opacity 1, pointer events on) only after a poll that ran while the slide showed, since the frame's last pagehide, and found the snapshot line with its commit, no font loading, and the four count cards with every ancestor at opacity 1 and no transform or the identity matrix.
2. If the deck can't read the frame's document (null, or a SecurityError), the still stays.
3. Every key pressed in the frame reaches the deck unless its target is a text field, and when the deck prevents its copy of the key, forwardKey prevents the original too.
4. An export prints the still and no frame.
5. Once shown, the frame stays shown when the slide leaves and comes back.

We assume the tracker starts its fade-in within 400 ms of the slide showing, before the deck's first poll. Neither engine's run found a shown sample with unsettled cards.

Tests

• CoverageEmbed.test.tsx (18 tests) enumerates 5 page texts × 2 font states × 7 card states, every combination of loaded × showing × document state, the reload, a cross-origin frame, and 8 keys × 7 targets.
• Each of eleven mutants of CoverageEmbed.tsx fails the suite (checks/mutation.txt). They skip the card check, the font check, opacity, transform or ancestors, accept any one card, or match the release without the commit. Others poll while the slide hides, keep the frame after a reload, show an unreadable frame, or let the year menu keep the arrows.
• bunx vitest run slideshows/appam-2026 passes 197 tests and analysis/test_finish_pdf.py passes 15, and tsc --noEmit and ESLint report nothing.

Browser runs

We ran the 31b59c3 preview in Playwright under Docker, routed under www.policyengine.org so the frame shares the deck's origin, with a 45-second dwell (rc14-reveal-*.log):

• WebKit and Chromium showed the frame about 1.3 s after the slide arrived, with all four cards at opacity 1 and no transform. No shown sample had unsettled cards (51 and 30 samples) or the fallback registry line.
• After a click in the search box, ArrowRight stayed in the box (step 13 stayed 13). After a click on the tracker's title, the arrows moved the deck (13 → 14 → 13). With the year menu focused, they moved the deck too (13 → 14 → 13), and the frame stayed shown.

Not verified here

• Firefox doesn't render the deck under routing. The production check after the merge covers Chromium, WebKit and Firefox.
• No PDF export ran through finish_pdf.py, whose off-address check would confirm the new rectangles.
• Nobody has run this head on the presenting laptop.

Reviews

• Two Opus delta reviews covered earlier heads: review-delta-609c612.md and review-delta-b3e057e.md.
• GPT-6.1 Sol reviews 4d3ec67 on two lenses (code and tests; claims, numbers and voice).
• The voice pass ran on all changed copy, against the whole voice guide.

🤖 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 6, 2026 12:51am UTC