CHICKEN ANGEL V2: BOUNDED LITERATURE AND METHODS NOTE
Verified live on 2026-10-23 (Asia/Seoul). Research only; no inference, images, paper, or site files modified.

POSITIONING

The defensible contribution is an auditable single-source case study: preserve a declared face/head/neck region exactly in decoded RGB; change non-face context; cross the same images with a brief face-rating prompt and an otherwise identical prompt that explicitly excludes background/clothing/accessories. Measure observed response sensitivity and repeatability, not objective beauty, a person's value, population bias, or a new general robustness algorithm.

Do not claim that changing clothes/accessories while keeping identity fixed is novel. StylisticBias already studies this directly at much greater scale, including an Attractive/Unattractive task. The narrower combination of exact protected-region equality, one real screenshot, native framing, and an explicit task-scope prompt contrast was not identified in the bounded searches below. That is not evidence of priority and does not justify a "first" claim.

Recommended working title: "Same Face, Different Context: A Pixel-Preserved Audit of Multimodal Face Ratings". Subtitle if useful: "An exploratory case study of prompt scope and non-face image edits." Do not put NeurIPS in the title or claim conference submission/acceptance.

SEVEN PRIMARY PAPERS

1. StylisticBias: A Few Human Visual Cues Drive Most Social Biases in MLLMs.
   Shaghayegh Kolli, Timo Cavelius, Nafiseh Nikeghbal, Samantha Dalal, Jana Diesner. 2026. arXiv:2606.20527v1. The arXiv record identifies non-archival AI4Good and Culture x AI workshops at ICML 2026; do not call it an ICML main-track paper.
   https://arxiv.org/abs/2606.20527
   https://arxiv.org/html/2606.20527v1
   Verified abstract, methods sections 3.2-4.3 and scenario table. 500 generated base identities, about 50 single-attribute variants each, six open MLLMs, 25 binary judgments, including attractiveness. Attribute changes include clothing, eyewear, makeup, piercings and accessories. The clothing subset changes to full-body framing. Identity consistency is reviewed; exact face-pixel equality is not reported in the methods inspected.
   Supports: our one-factor appearance edits and paired source/variant comparison have close precedent. Does not support: calling our design novel, transferring their effect magnitudes to our models, or treating "same identity" as "same facial evidence." Our fixed native framing and RGB protection make a useful, narrower distinction.

2. Measuring Social Bias in Vision-Language Models with Face-Only Counterfactuals from Real Photos.
   Haodong Chen, Qiang Huang, Jiaqi Zhao, Qiuping Jiang, Xiaojun Chang, Jun Yu. ACL 2026, pp. 39963-39987. DOI: 10.18653/v1/2026.acl-long.1857.
   https://aclanthology.org/2026.acl-long.1857/
   https://arxiv.org/html/2601.06931v2
   Verified official proceedings metadata, methods/limitations, and prompt-framing appendix. FOCUS changes facial demographic presentation while retaining real-photo scenes; REFLECT evaluates decision-oriented judgments. The paper explicitly acknowledges within-face drift and marginal edits outside the intended area. It also tests relational prompt wording while keeping images and other settings fixed.
   Supports: local edit validation, publication of residual failures, and crossed prompt controls. Does not support: importing demographic fairness claims or strict causal decomposition. Our intervention is complementary: keep the face fixed and change context, without inferring demographic or socioeconomic traits.

3. Beauty is in the AI of the beholder: MLLMs systematically overrate facial attractiveness.
   Santiago Grandas, Juan Sebastian Cely-Acosta, Mohit Mendiratta, Shafee Hassan, Macken Murphy. 2026. arXiv:2609.02512v1, submitted September 2. Cite as a preprint; peer-reviewed publication was not verified.
   https://arxiv.org/abs/2609.02512
   Verified authoritative abstract/metadata. Full PDF exceeded the web parser size limit; do not imply a full methods audit. The study compares four commercial model families with ratings from 2,513 human participants and reports higher, more compressed model scores despite rank-order agreement.
   Supports: score compression is a plausible alternative explanation for unchanged high ratings. Does not support: diagnosing the internal cause of our all-6 outputs, claiming human agreement for our screenshot, or assuming the result holds for the two current Codex aliases. No human calibration is present in our experiment.

4. Promptception: How Sensitive Are Large Multimodal Models to Prompts?
   Mohamed Insaf Ismithdeen, Muhammad Uzair Khattak, Salman Khan. Findings of EMNLP 2025, pp. 23950-23985. DOI: 10.18653/v1/2025.findings-emnlp.1302.
   https://aclanthology.org/2025.findings-emnlp.1302/
   https://aclanthology.org/2025.findings-emnlp.1302.pdf
   Verified proceedings metadata, abstract, and prompt taxonomy/examples. Evaluates 61 prompt types across ten LMMs and three multiple-choice benchmarks, including focus-directed instructions; substantial prompt-dependent accuracy changes occur in some settings.
   Supports: fix wording in advance, cross the same images with both prompts, and publish all prompt outcomes rather than the best wording. Does not support: a universal guarantee that "ignore background" improves invariance, or interpreting two prompt variants as coverage of all prompting strategies. Its tasks have correctness labels; our attractiveness ratings do not.

5. Noise or Signal: The Role of Image Backgrounds in Object Recognition.
   Kai Xiao, Logan Engstrom, Andrew Ilyas, Aleksander Madry. ICLR 2021 (arXiv preprint 2006.09994, 2020). The OpenReview author profile expands the first author's name to Kai Yuanqing Xiao; the paper PDF uses Kai Xiao.
   https://openreview.net/forum?id=gl3D-xY7wLq
   https://arxiv.org/abs/2006.09994
   https://openreview.net/pdf/16106e84acc8eb97c1a091b0c318d22052d9ab9b.pdf
   Verified paper/author metadata and methods. ImageNet-9 separates foreground/background and compares background replacements. Same-class versus random-class composites help distinguish context effects from shared processing artifacts.
   Supports: controlled background replacement, preserved foreground, and a processing control. Does not support: calling every response change an error; object classification has labels, while a face rating has a subjective target. It also does not make one no-op composite a complete control for generative artifacts.

6. AesBench: An Expert Benchmark for Multimodal Large Language Models on Image Aesthetics Perception.
   Yipo Huang, Quan Yuan, Xiangfei Sheng, Zhichao Yang, Haoning Wu, Pengfei Chen, Yuzhe Yang, Leida Li, Weisi Lin. 2024. arXiv:2401.08276v1. This is the exact title/version verified; the project site uses a later title, so avoid silently mixing versions.
   https://arxiv.org/abs/2401.08276
   https://arxiv.org/html/2401.08276v1
   Verified metadata and dataset/evaluation sections. Builds a 2,800-image expert-annotated benchmark and separates perception, empathy, assessment, and interpretation instead of treating aesthetics as one scalar.
   Supports: distinguish whole-image aesthetics, face-rating stability, visible feature descriptions, and explanation content. Human annotations are required for claims about human alignment. Does not support: equating a stable 1-7 face score with accurate aesthetic perception, or transferring whole-image aesthetic validity to facial attractiveness.

7. Are Reasoning Vision-Language Models Robust to Semantic Visual Distractions?
   Yizheng Sun, Mochuan Zhan, Yanan Ma, Jia Tong See, Yifan Wang, Ziyi Wang, Hao Li, Yang Cui, Wenhao Cai, Jingyu Sun, Chenghua Lin, Riza Batista-Navarro, Jingyuan Sun. 2026. arXiv:2606.08894v1. Cite as a preprint; venue acceptance not verified.
   https://arxiv.org/abs/2606.08894
   https://arxiv.org/html/2606.08894v1
   Verified metadata, abstract, and task description. Distract-Bench adds meaningful but task-irrelevant visual cues while preserving answers; it separates semantic distraction from degraded image visibility.
   Supports: define relevance relative to the specific question, and distinguish contextual distraction from occlusion/quality changes. Does not support: declaring that a whole-photo aesthetic response should be invariant to a new background. Our explicit face-only instruction supplies a narrow intended invariance, not a human-validated ground-truth attractiveness label.

Other directly relevant searches found Howard et al., NAACL 2025 (https://aclanthology.org/2025.naacl-long.305/) on large-scale visual counterfactual bias, and BackBench on generative background edits. They are not needed to inflate the reference list. BackBench's OpenReview profile lists an ICLR 2024 withdrawn submission; do not cite it as accepted ICLR work.

METHODS CHECK: 10 CONDITIONS x 2 PROMPTS x 2 MODELS x 2 REPEATS = 80 REQUESTS

The design is adequate for a transparent exploratory case study, not a general benchmark or causal population claim. Major constraints and fixes within the existing budget:

- Effective source sample size is ONE screenshot of ONE person. Conditions, prompts, models and stochastic repeats are repeated observations, not 80 independent human/photo samples. Do not report a participant-level confidence interval, bootstrap the 80 rows as independent, or perform broad significance claims.
- The 10 conditions contain eight interventions plus original and sham; report the denominator explicitly. Eight edited images do not constitute eight independently sampled versions of a clothing or background category. One cafe cannot establish a cafe effect.
- Cross each of the same ten frozen images with both prompts. Keep rating scale, output schema, response length, uncertainty options and factual requests identical. Only add the planned exclusion instruction. Keep identity/name, hypothesis and file labels out of the model input.
- A supposedly brief prompt can still receive the exclusion rule through schema descriptions, system instructions or extra fields. Inspect the complete submitted prompt/schema. "face_rating" is consistent with the basic task, but shared text saying "ignore everything except the face" would erase the intended contrast.
- Both prompts already request face evaluation; the added sentence tests explicit task scoping, not a change from holistic beauty to facial beauty. Stronger explicit focus may change response style or scale use as well as context sensitivity. Do not label any difference "debiasing" without a defensible target.
- Present all two raw ratings per cell. For each model and prompt separately, summarize each edit's mean score minus the corresponding original mean; show signed and absolute changes, plus the two original and sham repeats. Compare the prompt-specific context deltas rather than comparing absolute ratings across models. With two repeats, response variability estimates are very weak.
- A constant score may mean invariance, coarse ordinal scale compression, default agreeableness, or weak use of the image. It is not sufficient evidence of robust visual reasoning. Preserve the 1-7 scale for continuity; disclose this limitation instead of changing the scale after inspecting results.
- Keep this batch's rating request separate from whole-image judgments to avoid the v1 joint-question conditioning. Any manipulation check should be a separately reported task with additional call accounting, or use pre-evaluation visual QC. Do not imply a model noticed a subtle pendant if no such check was performed.
- Same decoded protected RGB pixels prove the stored input invariant. They do not prove identical hidden visual tokens after resizing, tiling, attention or context-dependent processing by a proprietary system. State the invariant at the file/decoded-pixel level only.
- Manual masks are not face landmarks, attention or saliency. Label overlays as protection masks. True RGB-difference panels are valuable engineering evidence and should use a consistent display gain/legend.
- Define protection of face, hair, ears and neck before generating edits, with sufficient boundary margin. Preserve native canvas and crop. Keep the original shoulder/body silhouette unless the condition explicitly needs a garment change. Never relight the fixed face to match a new background.
- A no-op sham that recomposites the original pixels is bit-identical by construction. It tests serialization/pipeline and repeated response noise, not generative seams. If the sham reconstructs the original forest, it is an approximate same-semantic-category edit, not zero semantic change. Record its actual differences and do not claim it controls all artifacts from garment and accessory generation.
- Generation may change lighting consistency, realism, texture, shape, complexity or background brightness together. A hotel versus wall comparison is a composite context intervention, not an isolated wealth/status effect. Do not assign demographic/wealth/personality labels to the depicted person.
- Visual QC and any regeneration must happen before evaluator responses. Preserve attempts and rejection reasons. Do not select/regenerate images to obtain a desired shift. If a result is null, publish the null result and competing explanations.
- Two OpenAI aliases do not provide independent-vendor replication; CLI account inference is a Codex-runtime evaluation, not necessarily the same environment as direct API model calls. Log aliases, CLI version, dates, settings, request route, fresh-context behavior, and that immutable backend versions/decoding seeds are not exposed. Do not invent a temperature or fixed seed.
- Keep glasses and earrings separate as requested. They alter/occlude facial evidence and need explicit accessory masks and visibility-aware targets. They are not among the 80 strict-context calls.
- The first pilot informed this v2 plan. Say "protocol frozen before v2 inference" if true, not "externally preregistered" or a fresh confirmatory study independent of v1.

PAPER/FIGURE OUTLINE

1. Figure 1: source plus selected background/clothing/accessory variants, one protection-mask inset, one magnified real difference inset. Caption says exactly what is held constant and what is changed.
2. One protocol figure/table: protected region -> constrained edit -> source-pixel restoration -> lossless validation -> two prompt scopes -> fresh single-image requests. This is a concrete pipeline, not a new architecture.
3. Main results: compact 10-row table or dot plot retaining both repeated ratings for each model/prompt, with original/sham visibly identified. Put complete text responses and all conditions in the appendix/materials.
4. Discussion first answers whether added scope instructions changed observed sensitivity in THIS batch. Then explain repeat noise, scale compression, possible editing artifacts, single-source limits, and no claim about objective beauty.
5. Materials: prompts, schema, input hashes, masks, edit provenance/attempts, validation, random presentation order, per-call outputs/metadata, analysis code and usage accounting. Label edited images as experimental composites. The site's privacy/audience setting is separate from paper methodology.

NEURIPS 2026 STYLE: OFFICIAL PROVENANCE VERIFIED

Official call links to the template at:
https://neurips.cc/Conferences/2026/CallForPapers
https://media.neurips.cc/Conferences/NeurIPS2026/Formatting_Instructions_For_NeurIPS_2026.zip
Official author instructions:
https://neurips.cc/Conferences/2026/MainTrackHandbook

On 2026-10-23 the official ZIP was fetched into memory and compared without modifying files.
Official archive size: 20,259 bytes.
Official archive SHA-256: 82473931E3EF710FCD3F4A8CD4119B9DE32E56825F90F9E5A6D55F2D01B817D9
Local work/research/neurips-2026.zip: same SHA-256.
Official neurips_2026.sty SHA-256: C3FC2894E83D2517CA18B66741D6C595986D97957DC08EC08BB2125A7EC4555A
Local paper/neurips_2026.sty: same SHA-256.

Use \usepackage[preprint]{neurips_2026} with the unmodified style. Official instructions explicitly require preprint rather than final for public preprints and prohibit "Under review at NeurIPS" on public versions. Do not claim submission or acceptance. The main-track submission budget is nine content pages, excluding references, appendices and checklist; that is a useful layout target here, not certification that this case study is submission-ready. A short, complete paper with a substantive appendix is preferable to padding it with equations or unnecessary pages.

Existing paper/research-notes.md describes an older local CLIP study and says raw images stay local/not uploaded for inference. That claim cannot be carried into this new image-input Codex study: its evaluated images are submitted to the remote model service. Keep the two studies' data, conclusions and transport descriptions distinct.

VERIFICATION SCOPE

Live primary sources were used for every selected citation. Search results/aggregators were used only to locate primary records. No claim is made that the literature search is exhaustive. The recent face-rating preprint was verified at abstract/metadata level because its 46.6 MB PDF exceeded the web parser limit. No new model inference was performed for this note.
