taste-score

name	taste-score
description	Score processed media on 5 aesthetic dimensions using Gemini Vision
trigger	pipeline
depends_on	["media-process"]
sandbox	true
inputs	[{"name":"batch_size","type":"int","description":"Max images to score per cycle","default":15},{"name":"model","type":"str","description":"Gemini model for scoring","source":"config.yaml#arcanea_claw.classify.model"}]
outputs	[{"name":"scored_count","type":"int","description":"Number of assets scored this cycle"},{"name":"hero_count","type":"int","description":"Number of assets that achieved hero tier (80+)"}]
dependencies	{"python":["google-generativeai","supabase","Pillow"]}
tables	["asset_metadata (UPDATE quality_score, quality_tier, status='scored')"]

taste-score

Sends processed images to Gemini Vision for a 5-dimension quality assessment:

1. Canon Alignment (0-20) — matches Arcanea fantasy aesthetic
2. Design Compliance (0-20) — composition, palette, professionalism
3. Emotional Impact (0-20) — evokes wonder, power, mystery
4. Technical Fit (0-20) — resolution, clarity, web-ready
5. Uniqueness (0-20) — distinct from generic AI art

Total score (0-100) maps to quality tiers:

• hero: 80+ (uploaded to Vercel Blob)
• gallery: 60-79 (uploaded to Supabase Storage)
• thumbnail: 40-59 (thumbnails only)
• reject: <40 (not published)

Processes at most 15 images per cycle.

Pipeline Position

Step 5 of 8 — runs after media-process, feeds media-upload and social-prep.