# Check the result without overstating the proof

Run these commands from the extracted archive with Python 3.10+:

```sh
python3 evals/check.py --package . --json
python3 evals/check.py --asset examples/chart.asset.json --root . --json
python3 -m unittest discover -s evals -p 'test_*.py'
```

The package check validates its inventory hashes, current assets, colour source,
resolved tokens, offline references and required files. Hashes detect accidental
change, not malicious replacement of the whole manifest. Pin the archive SHA
from your trusted handoff if authenticity matters.

The asset check validates the contract, safe local paths, permitted literal
colours, image alternatives, document language/title, duplicate IDs, ARIA ID
references and obvious unnamed buttons. For charts it compares the rendered
HTML table with supplied records. For the included interval encoding it also
checks range endpoints and diamond position against the numeric domain. It
rejects mismatched values, invalid intervals and unsupported encoding claims.
Other chart encodings need their own geometry evaluator before claiming numeric
rendering proof. Native code, PDF, Office files and raster pixels are not parsed;
they receive an explicit manual-review requirement.

Checks do not execute arbitrary application code, fetch URLs or install packages.
They do not measure the rendered contrast of photographs, check keyboard logic,
run assistive technology, prove semantics from pixels or award aesthetic quality.
Contracts are metadata, never a substitute for inspecting an artifact.

## Required rendered review

Mark each row pass/fail/not-tested with evidence (screenshot/export, dimensions,
state and observation). Do not silently convert not-tested into pass.

| Criterion | Pass condition |
| --- | --- |
| Identity | Neutral everyday Cut/current lockup, Current Blue tonal Facet at expressive scale, exact supplied bird geometry and clear space; historical unrelated colours are not current defaults. |
| Reading | Clear sentence-case hierarchy and readable Geist body/metadata/figures; no em dashes, eyebrows, overlines, monospaced styling or decorative numbering. |
| Photography | Relevant credible subject, natural editorial light/cool whites, no blue wash, wide/portrait/thumbnail crop review, uncropped master and provenance; generated concepts labelled and never documentary proof. |
| Iconography | Consistent existing Phosphor Regular, 20/24px optical size in 44px minimum targets, clear touch labels/accessibility names; no emoji, mixed families, decorative heading icons or redundant link arrows. |
| Component family | Brand/UI segments, tabs, buttons, fields and feedback use the same quiet rest/hover/press/selected hierarchy; fields have one 1px focused boundary, no outer halo; close glyphs avoid permanent white enclosures, overflow dots centre optically, nested radii follow the formula, and keyboard cues remain visible. |
| Cursor response | Only selected roots opt in; local low-opacity response, no idle/cursor replacement; exit/cancel reset, hidden/offscreen suspension and touch/reduced/solid/static modes checked. |
| Family resemblance | Cool off-white/charcoal structure, selective Current Blue, Geist-only presentation and exact spacing/geometry; no decorative border grids, outlined cards or compulsory image motif. |
| Data integrity | Units/source/timeframe, correct scale/labels/geometry, honest uncertainty, equivalent table and missing/zero/negative handling. |
| Interaction | Necessary actions discoverable on touch; rest/hover/press/release/focus/selection/disabled/pending/error/recovery checked; keyboard position preserved; repeated input interrupts older transitions. |
| Material | Bounded functional glass roots for buttons, menus, popovers, context panels and drawers; one owner per shared plane; 6px group inset and readable targets; actual mapped refraction, sharp labels and stronger hover/press splitting. No flat fill passed off as optical glass. |
| Choreography | Actual button press/release and menu/panel/drawer open/close/reverse exercised; GSAP/CSS ownership clear, fast input interrupts without queues or jumps, grouped content follows material, focus/keyboard/touch remain usable. |
| Optical fallback | Supported backdrop explicitly mapped; renderer/image failure, no-WebGL, solid and reduced effects stay usable; canvas/work bounds and hidden/offscreen suspension checked. |
| Accessibility | Actual composite contrast, 200% text/zoom, forced/solid/reduced modes, keyboard and assistive technology where applicable. |
| Layout | 320/390px, tablet and wide desktop for web; long labels and translated/expanded content; no obstruction or accidental overflow. |
| Digital studio | Report/view and presentation next/back/step controls work; figures retain readable sources/values/uncertainty across sizes. Optional attention light does not recolour encoded data or bloom over labels. |
| Identity study | Static default remains; optional Facet flight is finite, preserves the intact contour/planes, settles and supplies a still reduced state. |
| Production | Intended digital view inspected first; any secondary PDF/print export separately checked for fonts, links, pagination, image resolution and reproduction. |

The glass reference is a known-image artistic optical model, not native Apple
Liquid Glass or proof of arbitrary-DOM refraction. Inspect both a quiet rest
frame and hover/held-press/release in motion. Compare depth zero and normal
depth, prismatic on/off, fast selection changes and image/render failure.
Spectral highlights must never be the sole indicator of focus or state.

A source luminance bound is not proof of composite text contrast. Test labels
over the lightest/darkest supported backdrop and their actual selected/hovered
fills. Reduced motion must remove CSS and JS interpolation, not just one rule.
Treat GPU timing/power and physical-device feel as untested unless measured.

## Does the interface feel alive?

- Enter a real glass button: local light follows and the mapped image bends,
  rather than only repainting a flat fill. Hold press: lensing and colour split
  increase; release: the material settles without delaying the action.
- Alternate controls rapidly: no queued entrances, ghost selection, stale press
  state or return to an old pose. Reopen a closing surface from its current state.
- Open a real menu or context panel: the shell responds from its trigger, its
  actual actions work, labels remain crisp and keyboard/touch do not depend on hover.
- Open and dismiss a drawer: material, scrim and content move coherently; correct
  focus containment/return applies only when modal. Escape/outside behavior fits
  the declared contract and background tasks remain reachable when appropriate.
- Scroll/resize a supported backdrop: optical mapping stays aligned. Verify
  known-image scope rather than claiming generic DOM refraction.
- Enable reduced motion and solid mode during interaction: GSAP, CSS and shader
  interpolation stop together at a usable state. Closed/hidden roots stop drawing.

These are rendered interaction checks, not a prose definition of implementation.
Record which actual controls were exercised on the material-controls surface,
main navigation and inline examples; leave unavailable/unrun cases not-tested.

For a static document, interaction checks may be not-applicable with a reason.
For native apps, replace browser sizes with the actual platform/size classes.
For print, contrast in an RGB screenshot does not prove ink-on-stock legibility.

Any incorrect data, distorted logo, unavailable essential action or unreadable
required information blocks acceptance. Cosmetic improvements remain review
items. Record visual recommendations separately from confirmed functional bugs.
An `automated_pass` report always includes pending manual checks. Human review
and publication authority remain separate proof states.

### Reviewed finish checks

- Focus a text field by pointer and keyboard: one fine boundary and visible caret,
  without an outer halo. Buttons still show independent keyboard focus. Check the
  stronger high-contrast system treatment separately.
- Inspect cursor light over glass labels: soft diffuse light, with expressive
  refraction at the edge rather than a white bloom.
- Dismiss a menu/panel/drawer: optics stay present until it is hidden, including
  interrupted close/reopen. Edit a confirmed field and verify feedback no longer
  claims that the changed value is applied or kept.
- Check preference labels have at least a 44px target.
