# Beauty Design Compiler Aesthetic Forensic Audit V2
Audit date: July 27, 2026
Historical V1 public surface audited:
<https://cauldron-beauty-compiler.leightonrephoto.chatgpt.site/>
## Corrective implementation note
This document is intentionally retained as the forensic **before** audit. Its 64-page counts,
missing-gallery finding, typography limitation, and sampled V1 screenshots describe the published
baseline that motivated V2; they are not claims about the corrected evidence set.
The V2 implementation confirmed the central concern: categorical fingerprint uniqueness did not
guarantee perceptual distinctness or page-wide stylistic convergence. The corrective architecture
now requires exactly one art-direction capsule, capsule-owned coupled bundles, lineage validation,
salience/accent/cardification budgets, multiple deterministic candidates, independent quality and
distinctness evaluation, and a safe capsule fallback. The current V2 set has 128 accepted pages
and a committed 128/128 renderer-style migration. Its post-migration evidence has 512
exact-viewport Chrome screenshots and metrics, both intended font faces loaded in every record,
zero failed resources, 52/52 required four-viewport direct inspections passing, and 256/256
mobile/desktop axe-core runs passing the finalizer. The optional Responses screenshot critic
produced no accepted result because configured account controls prevented the path from running;
seven staged calls were discarded atomically. Final
release status is reported only by the V2 final QA summary and perceptual-similarity evidence; this
historical audit is not overwritten to make the baseline look better. No Lighthouse, field Core
Web Vitals, human-participant result, customer deployment, or current public URL is established by
the corrected local evidence.
## Overall verdict
The published V1 compiler has a strong art-directed baseline. Its four principal variants are
visibly different in type, geometry, palette, hero framing, and atmosphere; the first viewport is
consistently clear; calls to action remain prominent; and the sampled mobile renders reflow
cleanly without visible horizontal overflow.
The public matrix is not yet persuasive proof of “thousands of original designs.” It demonstrates
64 unique categorical design vectors, but only 20 distinct header-plus-hero class signatures,
reuses much of the same below-fold composition and copy, renders no gallery section at all, and
shows synthetic numbered fixture clones in 36 of 64 pages. The result reads as a polished
compiler QA matrix, not yet as a portfolio-grade showcase of customer-level originality.
No implementation was changed during this audit.
## Scope, capture, and evidence boundary
Nine clean public Chrome captures were created and directly opened for inspection:
| Step | Public state | Capture | Dimensions | General health |
| ---: | ------------------------------------------- | ------------------ | ------------------- | ------------------------------------------------- |
| 1 | QA index | `home-desktop.png` | 1440×1000 | Healthy, with evidence-navigation gaps |
| 2 | Atelier, `atelier-split-portrait--balanced` | desktop and mobile | 1440×6557; 390×7384 | Strong |
| 3 | Ritual, `ritual-quiet-column--balanced` | desktop and mobile | 1440×7238; 390×6881 | Visually strong; fixture quality weak |
| 4 | Studio, `studio-gallery-index--balanced` | desktop and mobile | 1440×6220; 390×6363 | Strong first viewport; gallery proof absent |
| 5 | Precision, `precision-toolline--balanced` | desktop and mobile | 1440×4067; 390×4950 | Clear and disciplined; fixture repetition obvious |
The nine captures are stored under `artifacts/beauty-public-audit-v2/`.
This V2 audit visually inspected those nine public screenshots. It did not visually inspect all 64
public pages. Algorithmic and DOM-level measurements cover the full 64-page set and are identified
as such. The previously published fidelity ledger reports a larger local inspection set, but that
claim was not substituted for direct visual evidence in this audit.
## 1. QA index
**Health: generally healthy.**
Strengths:
- The opening composition feels intentional and premium.
- Variant-colored status chips create quick visual grouping.
- Recipe name, variant, scenario, hero, typography, palette, and rhythm are exposed per card.
- The 64-page and 16-recipe claims are immediately understandable.
Risks:
- The oversized heading wraps `complete-page` as `complete-` on one line and `page` on the next.
The break looks accidental at 1440px.
- All cards use the same “Open complete render” action, so there is no quick way to compare
variants, filter scenarios, or identify the strongest showcase set.
- Architecture, fidelity, source, QA summary, hashes, and evidence limits are not linked.
- The page is a technical inventory rather than a review narrative. A new reviewer has to infer
what to inspect and what the fixture limitations mean.
Recommended correction:
- Keep this complete matrix, but add a top-level evidence navigation rail and an eight-page
curated showcase.
- Add filters for variant, recipe, scenario, hero, and rhythm.
- Add a visible legend distinguishing “structural link present” from “real external workflow
tested.”
- Prevent awkward hyphen-line wrapping in the title.
## 2. Atelier
### Desktop
### Mobile
**Health: strongest and closest to portfolio-ready.**
Confirmed strengths:
- Warm paper, oxblood accent, generous serif display type, and fine rules produce a coherent
editorial identity.
- The hero balances a long wordmark with a strong adjacent atmospheric image.
- Business type, Portland locality, phone, and booking actions are unambiguous in the first
viewport.
- The service hierarchy remains readable without card-grid clutter.
- Mobile moves to one focal image and preserves the sticky booking action without visible
horizontal overflow.
Visible limits:
- The full page relies on the same semantic sequence and many of the same section headings used by
the other variants.
- The long mobile capture reaches 7,384px. The content remains legible, but repeated generous
spacing becomes fatiguing.
- The intended Fraunces/Manrope token names are not authenticated by packaged font binaries in the
public evidence; platform fallbacks limit confidence in final typography metrics.
## 3. Ritual
### Desktop
### Mobile
**Health: strong art direction, weakened by test-data presentation.**
Confirmed strengths:
- Mineral surfaces, soft radius, rust accent, diffuse material image, and quiet serif hierarchy
make Ritual recognizably different from Atelier.
- The CTA remains high contrast despite the softer palette.
- Long-form service and guide content maintains a calm reading rhythm.
- Mobile reflow is deliberate rather than a simple scaled desktop grid.
Visible limits:
- Numbered repeated fixture names and services make the page visibly synthetic.
- The 7,238px desktop and 6,881px mobile page lengths amplify repeated content.
- The same “Verified perspectives,” “Plan your appointment,” and broad closing-slab pattern reduce
recipe-specific authorship below the fold.
## 4. Studio
### Desktop
### Mobile
**Health: strong first viewport; incomplete visual proof.**
Confirmed strengths:
- Uppercase graphic display type, vivid coral/red action color, tighter geometry, and polished
lacquer material make Studio immediately distinguishable.
- The service index is crisp and easy to scan.
- The layout feels energetic without becoming a collection of colored cards.
- Mobile maintains a clear hierarchy and usable sticky action.
Material failure in the evidence:
- This page is explicitly named `studio-gallery-index`, and its manifest selects a gallery
composition, but the page contains no rendered gallery section.
- The omission is factually safe: the renderer refuses to present generated atmosphere as
documentary customer work. However, this page cannot serve as visual evidence for the
`gallery-index` recipe.
The correct fix is not to weaken the evidence policy. Add development-only fictional gallery
assets with explicit documentary-fixture provenance and consent metadata, or publish a separate
decorative-atmosphere gallery recipe whose semantics do not imply customer results.
## 5. Precision
### Desktop
### Mobile
**Health: clear and disciplined; visibly stress-fixture driven.**
Confirmed strengths:
- Cool blue, strong rules, low radius, sans display type, and tool/material atmosphere produce a
distinctly precise character without aggressive grooming clichés.
- Telephone fallback is literal and prominent.
- The shortest sampled page still feels finished.
- Mobile preserves business identity, locality, service entry, and call action.
Visible limits:
- Repeated names such as `Haircut 03` and `Beard trim 04` read like database pressure-test output.
- The lower page relies on the same large closing color slab as the other variants.
- A compact Precision spacing profile could reduce vertical travel further without harming
clarity.
## 6. Full-matrix algorithmic evidence
The published `qa-summary.json` reports:
| Metric | Result |
| ------------------------------------------ | -----: |
| Complete pages | 64 |
| Master recipes | 16 |
| Principal variants | 4 |
| Full fingerprints | 64 |
| First-viewport fingerprints | 64 |
| Invalid records | 0 |
| Records with complete information coverage | 64 |
Each variant and each scenario appears 16 times. Each of the sixteen recipes appears four times.
Hero distribution:
| Hero | Count |
| ----------------------- | ----: |
| `type-led-atmosphere` | 15 |
| `central-monolith` | 13 |
| `split-portrait` | 11 |
| `offset-editorial` | 7 |
| `adjacent-landscape` | 5 |
| `vertical-gallery-lead` | 5 |
| `edge-image-rail` | 4 |
| `framed-panorama` | 4 |
Page-rhythm distribution:
| Rhythm | Count |
| ------------------------ | ----: |
| `quiet-progressive` | 16 |
| `precision-linear` | 13 |
| `editorial-alternating` | 11 |
| `indexed-modular` | 11 |
| `full-bleed-punctuation` | 9 |
| `gallery-led` | 4 |
All twelve configured typography pairing IDs appear, with four to seven uses each.
These are healthy categorical-distribution signals. They are not a substitute for perceptual
comparison.
## 7. Fingerprint forensics
The V1 uniqueness score is a weighted equality comparison over categorical manifest axes. It
correctly gives high weight to recipe, hero, typography, and rhythm and rejects exact and
first-viewport hash duplicates. It does not inspect rendered geometry, text density, image
composition, or screenshot similarity.
Across the 64 public pages:
- full structural class sequences are all unique;
- only 20 unique header-plus-hero class signatures exist;
- all 64 full vector fingerprints are unique;
- all 64 first-viewport vector fingerprints are unique.
This is not a contradiction. Color, typography, motif, palette, framing, and other axes can make
two first-view vectors unique while they share the same macro header and hero architecture.
Therefore:
> “64 unique fingerprints” means 64 unique resolved design vectors, not 64 perceptually unique
> pages.
The working tree contains an opt-in V2 semantic fingerprint and bounded tournament compiler, but
the public QA artifact identifies itself as V1. The public evidence must not use the presence of
V2 source as proof that V2 selection ran.
For the next public evidence set:
1. Render normalized screenshot regions with identical fixture content per recipe comparison.
2. Compute a perceptual hash plus structural/embedding similarity as a secondary signal.
3. Publish nearest-neighbor scores and a contact sheet.
4. Require visual review for candidates above a conservative similarity threshold.
5. Keep the manifest-vector fingerprint as the deterministic identity; do not replace it with
pixel hashing.
## 8. Gallery coverage failure
The 64 manifests select:
| Gallery composition | Manifests |
| ----------------------- | --------: |
| `dominant-plus-support` | 18 |
| `editorial-diptych` | 18 |
| `cinematic-rail` | 9 |
| `measured-triptych` | 7 |
| `focused-single` | 6 |
| `gallery-index` | 6 |
Rendered public gallery sections: **0 of 64**.
The underlying cause is inspectable:
1. The development Lab simulates a rich image count so gallery-gated recipes remain eligible.
2. It only binds documentary assets to the actual gallery.
3. The fixtures contain no eligible documentary work assets.
4. The renderer correctly omits a gallery with no real or allowed decorative asset.
This preserves the factual boundary and should remain. The QA generator must stop treating
simulated eligibility as visual coverage. Add separate assertions for:
- selected composition;
- component emitted;
- minimum eligible assets;
- layout-specific DOM shape;
- successful desktop and mobile crop;
- evidence provenance.
## 9. Fixture-copy quality
The stress-fixture utility repeats source records and appends two-digit suffixes after the source
inventory is exhausted. A rendered `<h3>` scan found:
- 36 of 64 pages with at least one suffixed synthetic name;
- 184 suffixed heading occurrences;
- 12 balanced pages affected;
- 8 founder pages affected;
- all 16 inventory pages affected.
Examples include:
- `Haircut 03`
- `Beard trim 04`
- `New-client consultation 04`
- `Custom facial 05`
- `Remy Hart 03`
- `Avery Moss 02`
This is acceptable for cardinality and overflow testing. It is not acceptable in a public
showcase intended to demonstrate premium customer presentation.
Separate the evidence into:
1. **Showcase fixtures:** carefully authored, unique fictional services, professionals, reviews,
policies, and locations.
2. **Stress fixtures:** deterministic repeated records, clearly marked as layout pressure tests
and kept out of the primary portfolio path.
## 10. Shared choreography and perceived originality
The shared semantic component system is the correct architecture. The current render layer,
however, hard-codes several prominent headings and broad section shapes:
- “Begin with what you need.”
- “Choose your professional.”
- “Verified perspectives.”
- “Plan your appointment.”
This is repeated across art directions, alongside similar guide and final-action sequencing.
Variant tokens make these sections attractive, but the repeated language and macro choreography
reduce the impression of a customer-specific design.
Improve variation without sacrificing factual or accessibility guarantees:
- register two to four bounded heading treatments per semantic section;
- let recipes select between open list, rule-led ledger, full-bleed punctuation, and editorial
split compositions;
- vary which middle sections share a surface or visual pause;
- allow recipe-specific label, index, and caption anatomy;
- preserve identical SiteSpec bindings, heading hierarchy, CTA semantics, and coverage reporting.
Do not use free-form model copy or arbitrary section order to solve this.
## 11. Historical V1 responsive, accessibility, and performance observations
### Confirmed from the screenshots and public DOM
- The sampled desktop and mobile pages show no visible horizontal overflow.
- Primary actions remain visually prominent.
- The sampled mobile pages show a persistent booking/call tray.
- Text and controls remain legible against the sampled palettes.
- Decorative imagery does not contain visible text or fake documentary evidence.
- The public HTML contains a skip link, one H1, explicit action links, and robots metadata.
### Risks and gaps
- Full-page mobile lengths range from 4,950px to 7,384px in the sampled set. Rich fixtures need a
compact mobile rhythm and stronger progressive disclosure.
- The WebP responses use the wrong MIME type, so current image success depends on browser
sniffing.
- Platform fallback fonts mean final line breaks and layout shift cannot be certified for the
intended font pairings.
- The public host does not apply the documented security-header map.
- A full keyboard path, screen-reader naming, reduced-motion behavior, zoom at 200/400%, and
high-contrast mode were not independently exercised in this audit.
- The published fidelity ledger reports axe, keyboard, and local timing results; this audit did
not rerun those checks against the public host.
Screenshots alone cannot establish WCAG 2.2 AA compliance or production Core Web Vitals.
## 12. Historical V1 prioritized design corrections
| Priority | Correction | Why it matters | Acceptance evidence |
| -------- | ----------------------------------------------------------------------------------------------------------- | ---------------------------------------------------------------------------------- | ------------------------------------------------------------------------------------------------------------- |
| P0 | Make every claimed component family visibly render in the QA matrix, starting with all six gallery layouts. | Current public evidence claims gallery variation without showing any gallery. | At least one desktop and mobile render per gallery family with eligible fixture provenance and DOM assertion. |
| P1 | Split polished showcase fixtures from stress fixtures. | Numbered cloned services and people visibly undermine trust in the design quality. | Zero synthetic suffixes in the public showcase; stress pages remain available under an explicit QA label. |
| P1 | Add perceptual nearest-neighbor evidence. | Vector uniqueness overstates visible uniqueness. | Public contact sheet plus normalized screenshot similarity, with thresholds and nearest matches. |
| P1 | Increase recipe-specific middle-page choreography. | First views differ more than full-page narrative and section language. | At least two bounded compositions per major semantic section and a reduced repeated-heading rate. |
| P2 | Package and verify approved self-hosted font binaries. | Typography pairing IDs are not yet visually authenticated. | Network/font audit, exact family match, no unexpected fallback, and no measurable font-driven CLS. |
| P2 | Add a compact mobile rhythm for rich/stress fixtures. | Sampled pages become 4,950–7,384px long. | No loss of critical information; reduced vertical travel and preserved 44px targets. |
| P2 | Turn the QA index into an evidence navigator. | The current grid is polished but technically opaque. | Filters, showcase path, direct evidence links, release version, and fixture limitations are visible. |
| P2 | Diversify final-action architecture within the registry. | Large closing color slabs recur across variants. | At least three semantically equivalent closing compositions across the 16 recipes. |
## 13. Successor evidence-set acceptance criteria recorded by the V1 audit
Before describing the public compiler as visually production-quality:
- visually inspect at least one desktop and mobile render for all 16 recipes;
- visually inspect at least 24 complete pages spanning showcase and stress conditions;
- render every gallery family at least once;
- remove synthetic suffixes from the showcase path;
- publish normalized perceptual-neighbor results;
- publish first-viewport and full-page contact sheets;
- package and verify intended fonts;
- verify the public host's MIME and security headers;
- rerun keyboard, reduced-motion, zoom, axe, link, and Lighthouse checks on the hosted artifact;
- publish the exact commit and QA artifact hashes.
## Evidence limits
- The visual findings are grounded in the nine screenshots embedded above and the public DOM
inspected during this audit.
- The 64-page distribution, class-signature, gallery, fixture-suffix, and link counts are
algorithmic observations over all published HTML and QA manifest files.
- No customer site, real booking provider, payment flow, protected Generator Lab session, or
production deployment was tested.
- No claim is made that all 64 historical V1 pages were visually inspected in this audit.
- No claim is made that screenshots alone prove accessibility compliance or user conversion
performance.