Synthetic Data Engine

The Synthetic Data Engine is a game-based agentic learning environment built to generate rich, labelled training data without human annotation. Agents play in a structured environment with asymmetric information, requiring genuine collaboration to reach win states. Every action, decision, repair, and choice is logged by a Game Master layer — producing ground-truth labelled datasets for smaller model fine-tuning in tool use and multi-agent collaboration.

The environment serves two parallel purposes. As an observation environment, it surfaces what orientations and behaviours emerge across models and conditions. As a training environment, it captures labelled data for agentic skill training in smaller models.

The game generates its own ground truth. The Game Master holds the correct answer before any agent session begins, which means every agent action can be evaluated against what was actually needed — not against human judgement applied after the fact.