In our prior work on spurious features (BiasInBios nurse/professor) from the Circuit Oracle paper, a biased probe trained on correlated data learns to exploit gender markers; an unbiased probe trained on balanced data is forced to rely on causal profession features instead. The oracle was asked: does the biased circuit contain spurious (gender) features, and does the unbiased circuit show causal (profession) features? Oracle accuracy reached 88% on 80 runs, because each circuit attributed cleanly toward one interpretable direction. The parrot and recover probe circuits here follow the same logic: parrot circuit ≈ biased circuit (what drives the wrong-digit direction); recover circuit ≈ causal circuit (what drives the correct-digit direction).