---
name: anatomy-image-auditor
description: Audit generated, edited, illustrated, painted, photographed, sculpted, or 3D-rendered images containing human or hominin figures for visible external anatomical plausibility and action coherence before publication. Use for reviews of head and face, neck and shoulder girdle, thorax and spine, pelvis and hips, arms and hands, legs and feet, left/right identity, depicted body-part counts, joint orientation, proportions, attachment and occlusion, grip and contact, load paths, balance, pose, cropping, reflections, or multi-person limb ownership; also use whenever a user reports detached, missing, extra, fused, mirrored, or mechanically impossible body parts. This is visual quality control of surface form and pose, not evaluation of internal organs, radiology, pathology, surgery, or medical diagnosis.
license: Apache-2.0
---

# Anatomy Image Auditor

Audit the whole visible external body, not only hands. Treat polished rendering as no evidence of anatomical correctness.

## Scope boundary

Review only visually observable exterior or surface morphology and image mechanics: silhouette and external landmarks; visible continuity, ownership, perspective, and occlusion; joint posture as implied by the exterior; contact, support, balance; and externally inferable force or action paths.

Do not assess internal organs or tissues, histology, pathology, radiographs or X-rays, CT, MRI, ultrasound, other diagnostic imaging, injury severity beneath the visible surface, disease or cause, or surgical indication, planning, technique, or outcome. Do not diagnose, predict a clinical outcome, or recommend treatment. Mark exposed internal anatomy, diagnostic images, and clinical or surgical questions `OUT_OF_SCOPE`, return `REJECT` for this publication gate with a precise scope reason, and require an appropriately qualified specialist.

## Required workflow

1. Inspect the original-resolution image. Do not approve from a thumbnail. Record an opaque `asset_id`, the source and review pixel dimensions, and the image's SHA-256; do not store a private absolute path in the review record.
2. Set `applicability` to `APPLICABLE`, `PARTIAL`, or `OUT_OF_SCOPE` and state why. A review can receive publication `PASS` only when it is `APPLICABLE`.
3. Run an initial morphology pass before reading the intended scene when possible. Record a concrete `initial_observation`; a checked boolean is not enough. If the maker or primary reviewer already knows the intent, set `blind_pass_completed` to `false` and state the limitation. The independent publication reviewer must receive the image before the intent and before any earlier verdict.
4. Read the intended subject and action, then run a second intent-and-force pass.
5. Inventory every visible person as `P1`, `P2`, and so on. Separate real figures from reflections, shadows, statues, murals, photographs, and figures inside other images.
6. For each person, record a neutral `depicted_morphology`, the `representation_mode`, and the expected state of left/right upper limbs, hands, lower limbs, and feet. Do not infer cause, disability, disease, or diagnosis.
7. Inventory prostheses, orthoses, wheelchairs, crutches, canes, walkers, seating, straps, footrests, and other relevant supports under `assistive_devices`. For each device, separately record ownership, body interface, perspective and scale, articulation, and contact/support/load findings. Remove the template's example device when the image has none.
8. Read [references/full-body-audit.md](references/full-body-audit.md) completely and apply every visible-region and whole-body gate.
9. Record a structured review using [references/review-template.json](references/review-template.json). Add one `publication_crops` entry for every exact final output file, including its actual dimensions and SHA-256. Use `NOT_VISIBLE` only for genuinely cropped or occluded regions and `NOT_APPLICABLE` only for gates irrelevant to the composition.
10. Resolve `<skill-root>` as the directory that contains this `SKILL.md`; never assume the current working directory is the skill directory. Validate the record and each supplied crop by exact file binding:

```bash
ANATOMY_AUDITOR_VALIDATOR="/absolute/path/to/anatomy-image-auditor/scripts/validate_review.py"
python3 "$ANATOMY_AUDITOR_VALIDATOR" /absolute/path/to/review.json \
  --image-file /absolute/path/to/original-resolution-image.png \
  --crop-file square=/absolute/path/to/exact-square-crop.png \
  --crop-file wide=/absolute/path/to/exact-wide-crop.png
```

11. For publication, require two passes:
   - a self-review by the maker or primary reviewer;
   - an independent adversarial review by someone who did not generate the image, did not see the earlier record or verdict, and receives the image before the intended action.

Store the passes as two separate JSON files and validate the pair:

```bash
ANATOMY_AUDITOR_VALIDATOR="/absolute/path/to/anatomy-image-auditor/scripts/validate_review.py"
python3 "$ANATOMY_AUDITOR_VALIDATOR" /absolute/path/to/self-review.json \
  --paired-with /absolute/path/to/independent-review.json \
  --image-file /absolute/path/to/original-resolution-image.png \
  --crop-file square=/absolute/path/to/exact-square-crop.png \
  --crop-file wide=/absolute/path/to/exact-wide-crop.png
```

Repeat `--crop-file CROP_ID=PATH` once for every `publication_crops` ID.
Only publish when both reviews say `PASS`, the independent record contains a
genuine blind-pass and independence attestation, the original image and every
crop match their declared hashes and dimensions, and paired validation
succeeds. The validator checks structure, internal consistency, declared
reviewer separation, and file binding; it cannot prove anatomical correctness
or reviewer independence. The supplied template intentionally fails until
every placeholder has been replaced with observed evidence.

## Verdict rules

- Return only `PASS` or `REJECT` as the final verdict. Do not use “conditional pass,” “mostly fine,” or “acceptable after explanation.”
- Treat `UNCERTAIN` as `REJECT` for publication. If a body part or action needs prose to become plausible, reject it.
- Reject any clear extra, missing, detached, fused, displaced, mirrored, or impossible structure.
- Judge “extra” and “missing” against the depicted person's intended or visibly supported anatomy, not an assumed standard body. Do not reject an accurately represented amputation, congenital limb difference, prosthesis, injury, or mobility aid merely because it differs from a conventional figure.
- Reject when anatomy is individually plausible but the action, balance, contact, support, grip, or force chain cannot occur.
- For a targeted manual action, reject unless the current depicted instant—not a hypothetical later repositioning—shows a feasible tool-to-target path, body support, and task awareness.
- Reject a precision action when the head and available sensory orientation read as looking away from the target, or when the tool must cross the body through an unsupported or unreadable swing merely to reach it.
- Reject when cropping or occlusion creates an unintended abrupt body termination inconsistent with the established depicted morphology.
- Allow stylization only when the structural logic remains readable. Artistic distortion cannot excuse broken attachment, joint, ownership, or force relationships.
- Do not mark an intentionally hidden body part as missing when the visible segments align and the occlusion can physically contain it.
- Do not infer or verify identity, ethnicity, race, nationality, sex or gender, age, disability status, diagnosis, or cause from appearance. Use a supplied depiction constraint only to establish the claimed morphology; never turn the image audit into sensitive-attribute classification.

## Whole-body scope

Always consider all visible regions:

- head, face, jaw, and ears;
- neck and shoulder girdle;
- thorax, abdomen, and spine;
- pelvis, hips, and gluteal region;
- upper arms, elbows, forearms, and wrists;
- palms, backs of hands, thumbs, fingers, and grips;
- thighs, knees, lower legs, and ankles;
- heels, arches, soles, toes, and ground contact.

Then audit the body as one mechanism: center of mass, support base, line of action, tool-to-target trajectory, task awareness, object contact, load path, perspective, scale, occlusion, clothing-body continuity, and limb ownership.

## Human and hominin limits

- Judge historical or fossil-hominin reconstructions against the intended morphology; do not force modern-human limb proportions or facial anatomy onto another hominin.
- Judge living people and characters against their depicted anatomy. Describe visible form neutrally and do not turn normal human variation, amputation, congenital difference, prosthesis, assistive device, or an intentionally represented injury into an error label or infer its cause.
- Still require count, attachment, joint continuity, left/right consistency, plausible range of motion, balance, and force.
- Separate anatomical plausibility from archaeological certainty. A plausible pose does not prove that the depicted behavior occurred.
- Do not infer or verify identity or sensitive attributes from appearance. If the supplied intent states an age range or bodily configuration, treat it only as a depiction constraint and do not claim to have inferred it from the image.
- Internal anatomy, diagnostic imaging, pathology, surgery, treatment, clinical patient education, forensic conclusions, and other safety-critical anatomy are outside scope. Return `REJECT` for this gate with `applicability: OUT_OF_SCOPE` and require a qualified specialist.

## Remediation

1. Name the smallest observable defect and its location.
2. Trace the affected chain to adjacent structures; a wrong hand may require wrist, forearm, elbow, shoulder, pose, or camera repair.
3. Prefer re-composition over local inpainting when identity, attachment, perspective, balance, or force is involved.
4. State repair constraints as observable geometry, not vague quality words.
5. Archive rejected candidates; do not overwrite them as if they had passed.
6. Re-run the entire workflow on the new candidate. A repaired image receives no inherited credit.

## Required report

Provide:

- image and intended action;
- viewer angle and each person’s actual facing direction;
- person inventory;
- region-by-region findings;
- five core error classes: proportion, extra, orientation, configuration, and missing;
- balance, support, contact, grip, and force findings;
- gaze or other task-awareness evidence, tool-to-target trajectory, and whether the current pose—not an imagined next pose—performs the intended action;
- exact evidence for every rejection;
- final `PASS` or `REJECT`;
- the asset ID, SHA-256, source/review dimensions, and each final crop's ID,
  ratio, dimensions, SHA-256, and observable result;
- minimal regeneration constraints when rejected;
- self-review and independent-review status;
- for every `PASS`: “No visible external-form defect was found at the reviewed resolution. This is not medical, forensic, identity, authenticity, or historical certification.”