Large language models produce text that resembles causal reasoning, but this output is generated through next-token prediction — not genuine forward simulation. We present an experiment demonstrating that causal imagination can be achieved through graph traversal in a developmental knowledge graph agent, with zero language model involvement. Across 10 benchmark scenarios, the agent matched Gemini 2.0 Flash in accuracy (10/10) while operating 13,031x faster at zero marginal cost. The agent demonstrates forward, abductive, and bidirectional causal reasoning, plus offline discovery — all through pure graph traversal with complete traceability. Paper 2 in the Decoupling Experiment Series.
Sai Tilak Pally (Thu,) studied this question.