We introduce W_Trans , a measure of model uncertainty extracted from the prediction interval width of a Feature Tokenizer Transformer trained on 35 economic-theme principal components derived from 212 anomaly signals. W Trans has zero unconditional return predictability (t ≈ 0.5) across all smoothing horizons and subperiods. However, the interaction of W Trans with the anomaly consensus signal µ generates Fama-MacBeth t-statistics exceeding 8 and a difference-indifferences portfolio alpha of +1.29% per month under Fama-French five-factor adjustment, with the sign-flip breakeven near the sample median of µ. The conditional signal survives microcap exclusion (t = 5.94), value weighting (t = 2.58), double-clustered standard errors (t = 6.18), a kitchensink specification with 35 theme controls (t = 5.74), and firm fixed-effect absorption (t = 2.37); a competing µ × σ interaction washes out completely (t =-0.19). The premium concentrates in low-attention stocks-low media coverage (t = 3.69) and low analyst following (t = 7.74)-and is absent among high-attention stocks (t < 1), consistent with a limits-to-arbitrage mechanism.