# Eve design repair · 8 September 2026

The user's examples exposed two shared authoring defects: comparisons always
said “Question” and the validator required final-row emphasis. All 49 original
comparisons were reviewed. Two became richer specimens; the remaining **47**
have authored three-field headings and an explicit focus rationale. Equal
alternatives use `null` emphasis. Four comparisons retain a justified focus,
including the original web/PWA outcome distinction. Every capability comparison
now has equal-weight rows.

Seven Eve slides now show containment, incoming graph relations, links from live
intent, uneven coverage, an illustrative work record, a queue admission gate and
walkthrough progress accounting. Their dated source scope remains visible. The
record and arithmetic are explicitly illustrative; no product screenshot or
measurement is fabricated. The repeated masthead is omitted on those specimens.
The 94 capability slides retain their detailed explanations, sources, review
questions and narration.

This repairs the reported defects and advances specific Eve explanations. It
**does not complete the previous whole-library design enhancements**. The
[per-guide review](../../authoring/eve-design-review.md) records the long table
runs and missing real-product examples. Complete typography, broader imagery,
full adoption and willing-reader evidence remain open in `TODOS.md`.

## Evidence and limits

- `before/compositions.json` records the baseline at `a60560f3aa`:
  578 unique slides / 601 guide instances. Of the 94 capability slides, 73 used
  tables or evidence comparisons. The after audit records the actual treatments;
  a layout count is not a design-quality score.
- `final-visuals/report.json`: **486** semantic/geometry checks covering all
  54 affected specimens at four standalone widths and expanded/collapsed desktop,
  compact and mobile center views. **144** captures; **six** defect controls
  reject forced emphasis, a changed heading, hidden/misattached graph relations,
  a missing refusal and a falsely scaled progress segment. Axe covers affected
  embedded specimens. Eight compact examples play real MP3 audio; controls
  remain reachable after scrolling. The test does not claim first-viewport fit
  inside every center frame.
- `capability-browser.json`: **282** exact authored-text/source/reflow checks
  across all 94 capability slides and three widths, plus four negative controls.
- `bank-layout.json`: one complete **578-slide** layout pass at desktop,
  compact and mobile, mobile source disclosure, and 578-page print verification.
  All **171** remaining tables preserve their exact fields and PDF header floors.
  This ran `--layout-only --draft`: its historical check labels mentioning
  accessibility do not mean that per-slide axe or every audio track was rerun.
- `center-smoke.json`: all **60** guide sequences, navigation, search, history,
  mobile focus/disclosure, narration interruption and interaction budgets pass.
  Detailed affected-slide center coverage comes from `final-visuals`, not this
  smoke run.
- `exports/report.json`: **21** changed guide PDFs / **167** pages regenerated
  and checked. `capability-pdfs.json` checks all 94 capability pages;
  `architecture-detail-pdfs.json` checks 50 older detail-guide page instances.
  The exact text checker accounts for Chromium painting positioned tree children
  and relation verbs after the folio in the content stream. Browser geometry
  separately verifies their visible attachment and direction.
- `logs/unit.log`: **113** focused Python tests pass. Scoped Ruff, ESLint,
  formatting, builder freshness, teaching crosswalk and assignment freshness
  also pass. Source assignment updates preserve every substantive teaching field;
  `assignment-review.json` records the six affected evidence links.

The final builder check verifies current narration fingerprints and each MP3's
bytes/hash. No narration was regenerated: the spoken material is unchanged.
Headless audio uses `--disable-audio-output`, the already-established Mac setup
that preserves actual decoding and media-clock progress without opening a
physical output device. Python checks use the existing Python 3.11 narration
environment; the system Python 3.9 cannot run the repository's newer PDF checks.

All expensive browser checks ran serially under `run-supervised.py` on the
16 GiB host. The bank run retained more than 6 GiB available. No Nx/app build,
model call or unrelated service test was required for these presentation changes.

## Optical review

The model inspected all twelve capability-guide sequence contact sheets and
the reported comparison, new diagrams, record and accounting example at actual
browser/PDF sizes. `sequences/` also provides the thirteen older Eve guide
sheets; those are available evidence, not a claim of a new complete optical
review. The seven `pdf-review/` images are page-fit 960px renders of the exact
exported PDFs. These retain readable body text, complete branches, arrow
endpoints, parentage and category labels in the model review.

`before/` retains the original graph/work openings and two rejected intermediate
designs. The initial relation arrows did not touch the subject; the final subject
boundary now receives each arrow. The initial record repeated its title and used
too much vertical space; the final specimen removes that repetition, shortens
annotations and gives the detailed explanation space in notes. Compact layout
also needed an explicit line height for its large progress total.

These are model observations and automated evidence, not a willing-reader study
or accepted visual benchmark. The user's original screenshot is the actual
reader feedback addressed by this repair. The full-size bank and capability
scratch captures remain local under `/private/tmp/eve-design-bank` and
`/private/tmp/eve-design-capability-check`; retained review captures are in this
receipt. `state.json` binds the final source, guide, PDF, report and retained
image bytes.

## Reproduction

Use the repository's dependencies, Python 3.11, Chrome and Poppler. Run browser
commands serially through `tools/presentations/run-supervised.py` and use fresh
output directories. The main scope-specific commands are:

```sh
python tools/presentations/build-eve-oshun.py --preview --check
node tools/presentations/tests/eve-design.mjs
node tools/presentations/tests/eve-depth.mjs
node tools/presentations/tests/eve-oshun-deck.mjs --layout-only --draft
node tools/presentations/tests/presentation-center.mjs --smoke
node tools/presentations/tests/eve-design-export.mjs
python tools/presentations/tests/check-eve-depth-pdfs.py --source-map authoring/eve-capability-guides.json --report /tmp/capability-pdfs.json
```

Set each command's documented output environment variable; the capability
semantic check uses `OSHUN_EVE_SOURCE_MAP=authoring/eve-capability-guides.json`.
Set `CHROME_CHANNEL=chrome` and `OSHUN_DISABLE_AUDIO_OUTPUT=1` for the shared
center/deck tests. The dedicated design matrix declares the same audio flag.
