A safe, bounded Computer-Use Agent (CUA) architecture for retro game emulation:

Visual Perception & Planning: Uses hosted multimodal Gemma to parse dual-screen DS layouts, perform OCR, and propose single-frame keyboard/touch actions.
Local Safety Guardrail: Leverages a local, text-only, fine-tuned Laya model to evaluate execution traces, detect state regressions, and guard against unauthorized actions or save overwrites.
Strict Execution Harness: Enforces window calibration, post-action semantic verification, and audit-grade JSON run tracing.