Evolving Strong Misreadings in the Residual Stream

We report an experiment at the seam of mechanistic interpretability and literary criticism. Using a base small language model (gemma-2-2b) and a sparse autoencoder (Gemma Scope) that decomposes its layer-12 residual stream into interpretable features, we isolate a 29-feature subspace encoding the motifs of a precursor text (the Kitab al-Tanazur) and run an evolutionary search over opening fragments whose machine-written continuations are scored for Bloomian misprision: a completion is rewarded when it stays saturated in the precursor's motifs yet binds them to internal company the precursor never gave them, while staying coherent and not quoting. The search reliably discovers a single rhetorical engine and drives the model's representation ~54x above control. A first run produces antithetical inversions (Bloom's clinamen); a second, steered toward deepening (tessera/apophrades), scores higher on the same metric, revealing that the objective is a swerve-magnitude detector agnostic to swerve-direction. A longform decay probe shows register stability is inversely proportional to swerve distance but indifferent to its direction; a winner-spread probe shows the negation and deepening winners occupy nearly the same representational region. We argue that strength of misreading is intensification, not distance, and that grading it as Bloom did would require an instrument for direction, not magnitude.Full HTML text: https://icra.tanazur.org/papers/strong-misreading/ — ICRA pre-print series: https://icra.tanazur.org/