Auto Validator

Auto-Validator Skill

Purpose

Automates Northcote Curio asset validation. Upload generated image → receive compliance JSON with scores, violations, and auto-generated correction prompt. Replaces 10-minute conversational validation with 30-second programmatic assessment.

Trigger Conditions

Use when:

• Gemini/DALL-E generates asset attempt
• Need compliance score (0-100 across 6 dimensions)
• Require iteration decision (≥90 package | <90 regenerate)
• Want correction prompt for next attempt

Validation Scorecard

Dimension 1: Geographic Authenticity (0-20)

• All specimens Australian endemic
• Test: "Did organism challenge European taxonomy?"
• Violations: Non-Australian fauna, generic specimens

Dimension 2: Translucency Physics (0-20)

• Light transmission (not glow) visible
• Internal structures shown through material
• Percentage compliance: 60-80% molt, 40-60% membrane, 20-40% leaves

Dimension 3: Scale Hierarchy (0-20)

• PRIMARY 1.5-2× SECONDARY
• SECONDARY 2-3× TERTIARY
• Clear focal points established

Dimension 4: Density Zones (0-20)

• Upper-left ≤20% coverage, 200×200px empty
• Lower-right ≤30% coverage, 150×150px empty
• Central 60-80% Wunderkammer density

Dimension 5: Background Color (0-10)

• Target: #1A1714 ±5% tolerance
• No sepia/brown drift
• Theatrical void maintained

Dimension 6: Typography (0-10)

• Serif font (Crimson Text style)
• Cream #F5F0E8 at 85% opacity
• 5-6 labels maximum
• Format: "Fig. X. Scientific name (Common)"

Workflow

Input: Image file path or upload Process:

1. Extract hex colors (sample 50 points)
2. Identify specimens (Vision API recognition)
3. Measure density zones (pixel coverage analysis)
4. Detect translucency (luminance gradient detection)
5. Count/validate typography (OCR)
6. Score each dimension
7. Generate violation list
8. Build correction prompt

Output: JSON structure

{
  "asset_id": "ASSET-3",
  "overall_score": 87,
  "decision": "REGENERATE | PACKAGE",
  "dimensions": {
    "geographic_authenticity": {"score": 18, "violations": []},
    "translucency_physics": {"score": 14, "violations": ["Spider molt opaque"]},
    "scale_hierarchy": {"score": 19, "violations": []},
    "density_zones": {"score": 16, "violations": ["Upper-left 25%"]},
    "background_color": {"score": 9, "violations": []},
    "typography": {"score": 8, "violations": ["7 labels (max 6)"]}
  },
  "correction_prompt": "CRITICAL FIXES:\n- Spider molt: Add '60-80% light-transmissive amber chitin'\n- Upper-left: Specify '200×200px COMPLETELY EMPTY'\n- Reduce annotations to 5 labels",
  "iteration_priority": "high"
}

Integration Points

With Flash-Sidekick:

• Call analyze_code_quality on generated prompt → identify vague language
• Call web_research_synthesis for specimen geographic validation

With Gemini:

• Auto-validator output → correction_prompt → paste directly into next generation

With Claude Desktop:

• Decision gate: score ≥90 triggers Asset-Packager skill
• Score <90 triggers Prompt-Composer with corrections

Usage Example

# Pseudo-workflow
result = auto_validator.validate(
    image_path="/downloads/asset-3-attempt-2.png",
    asset_id="ASSET-3",
    target_score=90
)

if result['decision'] == 'PACKAGE':
    asset_packager.run(result)
else:
    corrected_prompt = prompt_composer.apply_corrections(
        base_prompt=original_prompt,
        corrections=result['correction_prompt']
    )
    # Send to Gemini for regeneration

Efficiency Gain

Before: 10-15 min manual validation per attempt After: 30 sec programmatic validation Savings: 20× faster validation, 95% time reduction Scale Impact: 10 assets × 2-3 attempts = 3-5 hours saved

Implementation Notes

• Vision API for specimen identification + color extraction
• Pixel density analysis for zone coverage
• Luminance gradient detection for translucency validation
• OCR for typography verification
• Deterministic scoring (not subjective)

Replaces conversational validation with programmatic compliance measurement. Critical path acceleration for high-volume asset generation.