A Reproducible Framework for Structured Creativity in AI Localization and Cultural Design
© 2025 Hideyuki Okabe — MIT License
ModelRefiner v4.0 is the successor to ModelRefiner v3.5, extending its Refinement & Retraining loop with a Creative Induction Layer that fuses:
- HeartScape → Emotional & symbolic mapping layer
- SYNAPSE → Metaphor & perspective variation engine
Together, they enable reproducible creativity — where emotion, culture, and structure interact transparently within a verifiable QA loop.
Localization researchers, translation QA specialists, AI linguists, and creative QA developers working on:
- Generative content evaluation
- Multilingual cultural adaptation
- Brand tone alignment
- Structured creative rewriting
| File | Description |
|---|---|
docs/system_instructions.md |
Full technical specification (8K-level transparency) |
docs/translation_matrix.md |
Cross-locale mapping and creative axes |
docs/rope_metrics.md |
ΔS (Creative Entropy) & Creative Fit metrics |
docs/examples.md |
Practical use cases and sample data |
| Category | v3.5 | v4.0 |
|---|---|---|
| Loop | Evaluate → Refine → Retrain | Evaluate → Induce → Refine → Re-eval |
| Creative Layer | None | Creative Induction Layer (HeartScape × SYNAPSE) |
| Metrics | Fidelity, Structure, Style, Culture | ＋ ΔS, Creative Fit (Novelty × Utility × Brand Alignment) |
| Architecture | Single model | Dual-AI Protocol (Claude = emotion / ChatGPT = structure) |
| Domain Coverage | QA & Localization | ＋ Cultural Design, UX Text, Education, NPC Dialogue |
examples/in_lyrics.json— Lyrics localization (EN→JA)- (TBD) UX Microcopy — CTA tone & rhythm adaptation
- (TBD) Education — Pedagogic metaphor restraint
pip install -r requirements.txt python v4/scripts/pipeline_cli.py --in examples/in_lyrics.json --out out.jsonl The CLI is a mock implementation for demonstration. Replace induce() and refine() with actual LLM API calls using the provided prompts in v4/prompts/.
🧭 Notes Transparency-first: Every creative step is logged and reproducible.
Safety hooks: copyright, bias, and cultural constraints enforced.
Cultural intelligence: honors rhythm (ma/間), metaphor, and indirectness.
📜 License MIT License — see LICENSE
🪶 About ModelRefiner v4.0 extends the foundation of v3.5 into the realm of creative reproducibility — a framework where structure feels and emotion thinks. It aims to bring together fidelity, empathy, and brand-aligned creativity within a unified, auditable system for AI-driven localization and cultural QA.