# Beauty Design Compiler Fidelity Ledger
Review date: July 27, 2026
## V2 corrective evidence note
The reference comparison below records the original V1 review and is preserved as baseline
evidence. V2 replaces independent high-salience axis mixing with one capsule-owned visual lineage
and a bounded deterministic candidate tournament. Its current selected-page evidence comprises:
- 128 accepted fictional-fixture pages;
- 512 fresh browser screenshots at 390×844, 768×1024, 1440×1000, and 1920×1080;
- 52 directly inspected records covering every capsule and required boundary cohorts;
- 256 axe-core runs across mobile and desktop, with zero violations; and
- 32 controlled Chrome timing runs with zero runtime, resource, font, or layout-shift failures.
The renderer-style migration is committed for all 128 accepted pages. Every screenshot, runtime
record, axe run, and post-migration direct-inspection entry is bound to the committed migration
report and exact migrated HTML. All 512 runtime records loaded both intended font faces and
reported zero failed requests, broken links, console errors, page errors, resource errors, and
horizontal overflow.
The 70 axe `incomplete` contrast findings (221 nodes) were resolved independently from computed
browser styles; all resolved as passing. This is automated and Codex-inspected evidence, not a
human participant study. The controlled timing evidence is not a Lighthouse or field Core Web
Vitals result and does not provide LCP or INP where the available connector could not measure
them. The final public QA summary and perceptual-similarity report govern the V2 release result.
The optional Responses screenshot critic produced no accepted result. The path was unavailable
under the configured account controls; seven staged calls were discarded atomically, and no
partial result was admitted as evidence.
The V1 statements later in this ledger that self-hosted fonts and the asset URL resolver were
missing are historical. V2 browser capture waits for the final self-hosted fonts, records the
loaded files, rejects fallback-font rendering, resolves approved assets to browser URLs, and
verifies browser responses.
This ledger compares the fictional art-direction references under
`docs/templates/beauty-design-compiler/concepts/` with Chrome renders produced by the implemented
static renderer. The references establish composition, material, hierarchy, and restraint; they
are not customer sites or pixel-exact implementation specifications.
Local QA evidence lives under the intentionally ignored
`artifacts/beauty-design-compiler/` directory. In particular:
- `contact-sheets/reference-vs-chrome-desktop.jpg`
- `contact-sheets/reference-vs-chrome-mobile.jpg`
- `contact-sheets/desktop-{top,mid,bottom}-01..04.jpg`
- `contact-sheets/mobile-{top,mid,bottom}-01..04.jpg`
- `desktop-inspection.json`
- `mobile-inspection.json`
- `supplemental-responsive-inspection.json`
- `axe-results.json`
- `keyboard-inspection.json`
- `performance-results.json`
- `qa-summary.json`
## Reference-to-render findings
| Surface | Reference expectation | Chrome result | Visible mismatch | Fix or disposition |
| ----------------- | ------------------------------------------------------------------------------------------------------ | ---------------------------------------------------------------------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------------------ | ------------------------------------------------------------------------------------------------------------------------------------------------- |
| Atelier desktop | Warm editorial paper, restrained oxblood, asymmetrical portrait/media split, fine rules, clear booking | Warm canvas, oxblood action, large editorial wordmark, strong adjacent atmospheric image, visible booking and service actions | Reference has denser magazine-like section choreography and authentic editorial portraits; render is calmer and uses one decorative material study | Intentional evidence-safe deviation: no fake people or results. The shared rhythm preserves the editorial hierarchy without copying the reference |
| Ritual desktop | Quiet mineral palette, generous whitespace, soft asymmetry, booking never hidden | Centered serif hierarchy, low-noise mineral surface, restrained rust action, large atmospheric mineral image | Reference begins with a tighter split hero; selected recipe resolves to a more spacious centered composition | Accepted recipe variation. The reference's calm tactility and booking clarity are retained |
| Studio desktop | Graphic energy, strong type, one vivid accent, modular service rhythm | Oversized uppercase type, crisp grid, cobalt action, polished lacquer/glass visual, indexed next-section transition | Reference example is predominantly dark with lacquer red; the generated candidate selected a light approved Studio palette and cobalt accent | Intentional seeded variation, not a fidelity defect. Graphic hierarchy and one-accent discipline match the design thesis |
| Precision desktop | Exact alignment, condensed type, metal/tool material, strong technical metadata, telephone fallback | Tight axis, lower-radius geometry, cool-blue action, large aligned wordmark, metal/comb atmosphere, explicit “Call the studio” fallback | Reference uses a denser black/white editorial sheet; the historical V1 render was more open and used platform typography | V2 closed the font-packaging gap; layout density remains an intentional recipe expression |
| Mobile 390×844 | Immediate identity, category/locality, one clear action, no desktop-only navigation, safe-area tray | Identity, category/locality, call action, service action, one focal media moment, and sticky call tray are visible without horizontal overflow | Reference shows authentic brow imagery and a denser information sequence; tested render uses a non-documentary neutral atmosphere and a simpler first screen | Intentional evidence boundary and mobile simplification. Essential facts and actions remain available below the first screen |
## Material issues found and resolved
| Finding | Severity | Resolution | Evidence |
| ---------------------------------------------------------------------------------------------------------------------------- | -------- | -------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------- | --------------------------------------------------------------------- |
| A SiteSpec with an empty repeatable-service collection could omit approved service bindings | High | Renderer now falls back to the approved service binding before coverage validation | Renderer unit coverage and all 64 QA records report complete coverage |
| Testimonial strings without explicit verification metadata could be mistaken for proof | High | Proof rendering now requires `verified: true`; unverified entries are intentionally omitted | Renderer tests and no-proof/minimal-proof QA fixtures |
| Fixture atmosphere files needed a stronger boundary from real business evidence | High | Fixture media require both `allowDevelopmentFixtureMedia: true` and isolated fixture compiler provenance; metadata remains `aiGenerated: true`, `role: "decorative"`, `documentaryEvidence: false` | Fixture media tests and six visually reviewed WebP assets |
| Rich-image controls in the Generator Lab could affect recipe eligibility without carrying the accepted candidate fingerprint | Medium | The Lab simulates approved content shape only for compilation and retargets fixture-only request/provenance hashes to the compiled fingerprint before rendering | Generator Lab focused tests |
| Full-page Chrome capture exceeded the browser timeout | Low | Each of 24 complete desktop pages and each mobile recipe was inspected through top, middle, and bottom viewport captures plus bounded DOM checks | Desktop/mobile inspection JSON and contact sheets |
## Responsive and interaction fidelity
- Desktop: 24 representative complete pages were visually inspected at 1440×1000, with top,
middle, and bottom captures. All four variants, all 16 recipes, and rich, founder, inventory, and
minimal-proof conditions were represented.
- Mobile: all 16 recipes were inspected at 390×844 with top, middle, and bottom captures.
- Supplemental widths: `ritual-asymmetric-mineral--balanced` passed at 768×1024 and
`studio-gallery-index--balanced` passed at 1920×1080.
- Every inspected page had one H1, no heading jump, no duplicate ID, no broken image, no horizontal
overflow, no empty conditional section, no dead primary action, and a primary action in the
first viewport.
- Every mobile recipe had a visible sticky booking tray, a valid tray target, and no undersized
action or text-link target in the automated check.
- Keyboard inspection confirmed a visible 3px focus outline, skip-link-first order, and reachable
wordmark, navigation, header action, hero action, and service action.
- Sixteen axe-core scans using WCAG 2 A/AA, 2.1 A/AA, and 2.2 AA tags completed with zero
violations and zero incomplete checks.
## Performance evidence boundary
The 16 recipe representatives were loaded from a local static server. DOMContentLoaded measured
163.2–195.0ms and load completion measured 166.2–197.8ms. The pages contained 29,322–31,743
decoded HTML bytes, one image resource, no external scripts, and measured CLS of 0.
These numbers are useful regression smoke evidence only. They are not Lighthouse results, do not
measure a production network, and do not substantiate production LCP or INP. Hosted Lighthouse,
real-device, and real-user monitoring remain launch checks.
## Intentional deviations and open fidelity limits
- Generated fixture atmospheres are deliberately non-documentary. They do not imitate real staff,
customers, results, interiors, storefronts, or products shown in concept references.
- The compiler produces a curated family, not a pixel clone. Palette, hero, rhythm, and typography
may differ when the accepted deterministic seed selects another recipe-compatible option.
- V2 fixture evidence uses packaged self-hosted font binaries and retains fallback stacks only as
resilience. Both intended faces loaded in all 512 post-migration measurements.
- Approved fixture assets resolve to browser URLs and produced zero failed resource requests in
the 512 post-migration measurements. Connecting provider-owned production storage to the
production R2/public-URL path remains a separate deployment concern.
- The protected Generator Lab must be exercised with a real operator session; QA artifacts do not
bypass its authorization.
- No customer production deployment was performed, and no outreach, billing, DNS, domain, hosting,
or production control was changed.