# Models, quality and cost · chapter 11 · verification receipt · 10 September 2026

Scope: the fifth chapter of the _How Eve works_ track, rewritten in the Eve
edition — seventeen slides moved out of four shared manuscripts into one, five
chrome slides added, and every binding re-reviewed. The
[brief](../../authoring/eve-quality-chapter.md) records what the chapter is for
and why each treatment was chosen; [state.json](state.json) holds the exact
counts, durations and tool results.

## What was checked

- **Build freshness.** `build-eve-oshun.py --check` passes on the delivered
  tree: 1,416 unique slides, 658 sources, 75 guides. The chapter is 22 slides in
  three sections of five, six and six, running 30.6 minutes.
- **The stat panel is re-derived, not transcribed.**
  `eve-model-capability-floor` counts the registry's nine legs by status — four
  bound, two operator-configured, three not admitted — and a new test in
  `test_build.py` parses the leg list and each status count out of the pinned
  `model-registry.ts`. A leg added, removed or rebound in the source fails the
  suite rather than ageing on the slide.
- **The cost numbers are not a stat panel, on purpose.** The plan suggested one
  for cost and latency. Those figures are invented units chosen to show
  arithmetic, and every stat-panel value has to cite the source it was read from
  — so a panel there would have been a false citation. They stayed a compare
  panel and a card grid, with the notes saying plainly that they are
  illustrative.
- **One table, as planned.** `eve-model-registry`: four legs, their pinned
  defaults and what each pin means, side by side. Nothing else in the chapter
  needed one.
- **Assignments re-reviewed, not re-pointed.** Five claims across three
  assignment files bind slides this chapter rewrote — floors and release gates,
  the prompt-change recipe, escalation, judge validation and cost and
  performance. Each was read against its new slide before the fingerprint moved,
  and each carries a dated note.
- **Narration.** All 22 tracks rendered and verified; the library-wide `--check`
  verifies 1,416 tracks (1,847.7 minutes).
- **Browser rendering.** `eve-edition-review.mjs --guide eve-quality` sweeps all
  22 slides at 1440×900, 1280×720 with audio chrome, 390×844 and in print with
  no findings — on the first attempt, for the whole chapter.
- **PDF.** All 22 pages pass the exact-text contract.
- **Suites and lint.** Python 456 passed; Node 443 of 445; Ruff clean; Prettier
  clean on everything changed. No JavaScript changed, so ESLint had nothing new
  to read.

## What was not checked

- No human reader session and no listening review at normal speed.
- The narration was rendered on this host's CPU rather than the Apple GPU the
  rest of the library used; only the MP3 bytes differ.
- No capture was taken. `assets/eve-product/` holds nothing showing a model, a
  deck run or a cost report, so the treatment checkbox requiring one real
  capture per chapter stays open.
- The illustrative cost arithmetic is not a measurement and is not presented as
  one; no provider was called while writing this chapter.
- `test_v2_rollback_session` still fails against its published baseline;
  re-baselining it is a decision for the user.
