Background, Clothing, and Multimodal Face Ratings
Companion materials (23 October 2026)

Editable diagrams: paper/context-audit/design/figure-designs.drawio contains five
diagrams.net pages. Matching SVG files are in paper/context-audit/figures/.
Text, arrows, charts and photographs are separate objects. The perspective image
planes are display-only transformations; experimental input files are unchanged.

Unzip all six archives into the same directory. They share one top-level
chicken-angel-context-audit folder and contain non-overlapping data files.

records-and-source.zip: prospective protocol/freeze, prompts, schema, masks,
80 complete answers with event logs and request metadata, aggregate CSV/JSON,
secondary analysis, image-edit provenance, paper source, figures and PDF.
This now includes post-hoc spatial measurements and recorded local landmark
coordinates; these are not VLM attention or attribution maps.
inputs-a.zip: original, processing sham, three background edits (native PNG).
inputs-b.zip: three garment edits and two accessories (native PNG).
generation-candidates.zip: generated candidates, including the rejected first
hotel attempt. These are not the final pixel-preserving evaluator inputs.
pilot-inputs.zip: all nine native inputs from the earlier three-photo pilot.
pilot-records.zip: its 36 retained responses, raw events, masks, protocol,
freeze and code. Two recorded connectivity preflights remain excluded.

The display previews on the site are resized JPEGs. Only the native PNGs in
inputs-a/b.zip and pilot-inputs.zip are the evaluated files and subject to
exact pixel guarantees.
Each archive contains a manifest of SHA-256 hashes. The prospective freeze
predates evaluator calls. validation.json was produced before visual review;
its old quality_review="pending" fields are superseded by quality-review.json,
which was also frozen before inference. publication-audit.json is an independent
post-evaluation verification of the canonical pixels and original file hashes.

To reanalyze, install Python with numpy, Pillow and matplotlib and run:
  python paper/context-audit/build_evidence.py
Compile paper/context-audit/main.tex with Tectonic or the Codex LaTeX editor.
The earlier pilot is shown in the main paper and remains a separate protocol.
There are four source photographs, 19 input files and 116 retained VLM
responses. The phases are not pooled into a single effect estimate.

Landmark diagnostics were computed afterwards on CPU with MediaPipe 1.0.1.
Stored analysis/landmarks.json includes all coordinates, settings, image hashes
and the official model-bundle SHA-256. To recompute that diagnostic, run
  python paper/context-audit/analyze_landmarks.py
The script downloads the official model bundle if it is missing. No remote
VLM call is made. Other reanalysis uses numpy, Pillow, matplotlib and scipy.

Frozen source scripts retain local paths for provenance. The stored masks,
inputs and answers are authoritative; regeneration is not required. A live
evaluation needs a compatible Codex installation and existing account access,
consumes usage, and may change as model aliases change. Do not overwrite the
preserved run. No model is called by the static site or the evidence builder.

This is a four-photograph, one-person case study, not a human beauty benchmark. The user
provided the screenshot; a source-video URL and redistribution rights were not
established. The companion materials are publicly accessible; the images are
not an openly licensed face dataset. The paper is an independent preprint using the official unchanged
NeurIPS 2026 style in preprint mode; it is not submitted, reviewed or accepted.
