Low-Cost Geometric Directions in Frozen Speech Models: A Deepfake Hardness Signal

Deepfake-speech detectors are inconsistent in two systematic ways: against a fixed detector some synthesis systems evade detection far more often than others, and a detector trained on one recording domain often fails on another. Working at small scale---one GPU, a frozen $\sim$95M-parameter encoder, no detector retraining---we show a meaningful fraction of this inconsistency is legible in the representation's \emph{geometry} ($R^2{=}0.277$, $\approx$32\% of the reliability-corrected ceiling). Inside the encoder is a single natural-versus-synthetic direction; a system's position and spread along it track detection hardness, recurring within every corpus we test though \emph{which} statistic carries it differs---spread ($\mathrm{sd}_{\mathrm{along}}$) on the 61-system MLAAD, where it is the strongest of 83 quantities and the only one to survive a global false-discovery-rate correction; position on the few-system ASVspoof/In-the-Wild corpora (a pre-registered cross-corpus transfer was inconclusive, $\rho{=}0.25$, $p{=}0.41$, $n{=}13$). The direction rotates across datasets to near-orthogonality, showing a geometric correlate of transfer failure, and serves as a detector-training-free corrective signal conditional on headroom: it cuts a deliberately undertrained detector's error by $\approx$21\% on ASVSpoof2019 ($\approx$40\% in-dataset on MLAAD) and repairs a \emph{naturally} weak cross-domain detector (AASIST) by $26$--$33$ EER points, leaving an already-strong one unchanged.