AI is running out of real data. Synthetic data is the backup plan. But most of it has a logic problem. Labs now train models on data written by other AI. It's cheap, fast and private. However, there's a massive flaw. AI writes what sounds likely, not what's correct. It breaks basic rules. It repeats common examples and skips the rare ones. The next models learn those mistakes. CircuitSynth, from researchers at the Department of Computer Science @UniofOxford, @flock_io, and the Institute of Logic and Computation, TU Wien, splits the job in two. First, a small logic system picks the facts for each example. It can't pick combos that break the rules. A "pregnant male patient" is impossible by design. Then the AI turns those facts into sentences. It can't add facts of its own. The test was hard logic puzzles. @AIatMeta's Llama 3.1 8B followed all the rules 12.4% of the time. CircuitSynth hit 100%. It also covered far more rare cases, 52% against 35% for the next best method. That's the gap between messy data and data you can train on. The catch is that it struggles when the rules get very complex. Clean rules in. Clean model out.