Audit Before Design: Failure Modes of SAE-Guided Steering in Genomic Foundation Models

Scientific foundation models are increasingly used to generate candidate biological artifacts, but apparent controllability can be misleading when proxy metrics reward artifacts rather than function. We present a diagnostic audit of SAE-guided activation steering in Evo 2, testing whether interpretable sparse autoencoder features can be treated as reliable control knobs for genomic sequence generation. Across an exploratory 62-feature layer-26 screen, output effects are highly heterogeneous: one feature behaves as a collapse-prone low-complexity direction, another acts as a reproducible G/T-rich steering direction, and most tested features have weak descriptive effect scores. Mechanism controls show that the collapse-prone direction fails under both SAE clamping and direct decoder-direction addition, indicating that the failure is not merely an encode-decode projection artifact. Natural activation analysis over 2,000 human chromosome-22 windows grounds focal features before intervention, showing that collapse can arise from out-of-distribution amplification rather than from simply copying naturally collapsed sequences. We further evaluate generated human-context inserts with frozen Enformer and Borzoi regulatory predictors, a generation-cost-only Enformer reranking proxy, and a small multi-feature steering stress test. The key result is that a composition-matched Enformer shuffle control recapitulates the apparent central-activity gain, showing that the strongest biological-looking signal is largely composition-driven rather than evidence of preserved regulatory grammar. These results support an audit-before-design view of scientific activation steering: interpretable latents can be useful causal handles, but scientific control requires mechanism checks, sequence-quality tests, natural activation grounding, independent functional predictors, composition-matched controls, and multi-feature interaction audits before generated biological artifacts should be treated as design candidates.