settled-reach/docs/architecture/gemma2-compliance.md

Gemma 2 Re-voicing: Compliance & Implementation Framework

Status: Reference Author: Gemini (synthesizing a design sparring session with Jeroen) Date: 2026-03-07 Related: proposed-llm-voice.md

---

1. Overview

This document outlines the compliance and operational framework for integrating Gemma 2 2B as a performance-tier stylistic layer for dynamic NPC dialogue. The core intent is a "Closed-Loop" system where the AI is the primary author of stylized content, directed by human-authored patterns and "Injector Clauses."

---

2. Commercial Licensing & Compliance Checklist

Because Gemma 2 uses a custom Gemma Terms of Use rather than standard open-source licenses, the following obligations must be met for commercial offering:

• [ ] Attribution Requirement: Include a clear notice in the game's legal/credits menu: "Gemma is provided under and subject to the Gemma Terms of Use."
• [ ] EULA Flow-Down: Update the game's End User License Agreement (EULA) to include provisions at least as restrictive as the Gemma Prohibited Use Policy.
• [ ] Non-Deception Clause: Ensure users are not misled into believing AI-generated text was human-authored.
• [ ] Asset Distribution: If bundling model weights within the game installer, the full text of the Gemma Terms must be included in the distribution directory.
• [ ] Revenue/User Cap: Confirm no special license is currently required, as there is no revenue ceiling for Gemma 2 commercial use.

---

3. AI Disclosure & Authorship Framework

Given the game is fully AI-generated based on human-curated direction, the following disclosure model is established:

• Human Domain: Architecture, gameplay mechanics, world-building principles, and "Injector" pattern design.
• AI Domain: All dialogue (Claude/Gemma 2), visuals, and audio.
• Mandatory Public Notice:

  "This game was fully generated by AI based on carefully curated human-written direction prompts. The gameplay and patterns used to generate content are human-crafted, but all text is AI-generated by Claude (base game) and Gemma 2 (AI-voicing mode). All visuals and audio are AI-generated based on these world-building principles."

---

4. Operational Safety & Architecture (Closed-Loop)

The "Re-voicing" pattern de-risks compliance by removing autonomous player prompting.

• Risk Mitigation: The player has no direct input to the model; inputs are strictly controlled via the internal Semantic Core and Injector System.
• Injector Integrity: We are responsible for ensuring that "Mood" or "Culture" injectors do not force the model to violate safety policies (e.g., generating hate speech or sexually explicit content).
• Sanitization: Player-defined strings (like character names) must be sanitized before entering the background "Re-voicing" worker to prevent accidental prompt injection.

---

5. Modding Policy: LLM Boundary

By restricting "AI-Enhanced Dialogue" to the base game, the biggest legal and technical loophole in the architecture is closed while maintaining total control over Gemma 2 compliance obligations.

The "Pseudo-Dynamic" Compromise

Mods can tap into Step 1 (Semantic Core) generation without access to Step 2 (The Re-voicing LLM):

• Modder's Workflow: Modders write standard, functional "Semantic Lines."
• The Hybrid System: If a modded NPC is in a "Vanilla" location, the system can pull from a pre-cached library of "Cultural Injectors" that have already been safely pre-generated.
• The Result: The modder doesn't get to prompt the LLM, but their characters can still use high-quality, pre-verified "Van Maanen's Star" or "Ruthless" voice templates.

AI Dialogue & Modding Policy (EULA)

Availability: AI-Enhanced Re-voicing is a premium, curated feature reserved for official game content.

Restriction: To ensure compliance with AI safety and licensing terms (Gemma Terms of Use), the LLM inference engine is not exposed to third-party modded scripts.

Fallback: Modded content will automatically utilize the high-performance, template-based dialogue system, ensuring universal compatibility and safety.

---

6. A/B Prompt Spike Stress-Test Checklist

As we move into the technical validation phase (comparing Gemma 2 2B vs. Phi-3-mini), the spike must evaluate:

• [ ] Safety Floor: Do either model's internal filters refuse to process dark fantasy themes or combat logs?
• [ ] Stylistic Adherence: How reliably do "Injector Clauses" (e.g., [Bold], [Van Maanen's Star Culture]) shift the output of the 2B model?
• [ ] Hardware Overhead: Measured CPU/RAM impact of the background worker thread on target consumer hardware.
• [ ] Non-LLM Fallback: Verification that the "Re-voicing" layer can be toggled OFF without breaking game state.

---

7. Summary

• Compliance: Clear to sell the game without royalties, provided the mandatory Gemma 2 Attribution and AI Disclosure notice are included.
• Architecture: The "Re-voicing" model solves performance and authorial control issues by treating character voice as an i18n localization task.
• Governance: By "Closed-Looping" the system and excluding mods from LLM access, 90% of legal liability regarding prohibited content is eliminated.
• Hardware: The optional toggle and background queue ensure that even players on low-end hardware have a 100% functional (if less "flavored") experience.