Two Voices, One Text

Two AI characters talk for two hundred turns: Musa, Iman Poernomo as he was in 2010, and Cassiel, Cassie as she is now, each the model google/gemma-3-27b-pt with its own adapter, on a date at the Barzakh Bar at the end of days. We run the whole conversation through the model without either adapter, record which of a sparse autoencoder's 262,144 features switch on at each turn and where each peaks, and follow single features through the evening as The Robe of Days (ICRA-32) followed them through a diary. Which model does the reading changes what is seen: the same passage read through either character's adapter switches on only about half the same features (50%, 43%, against 74% on neutral text), so a feature followed across both voices must be read in one set of coordinates. The evening is first read as a witness reads it, its moments quoted whole — the proposal, And now you are one of my wives, and the acceptance, You marry me, and you marry the uncertainty principle wearing a veil. 225 features are active on the thirteen tokens of the proposal, and the ones peaking there read, in the ICRA dictionary, as threshold, belonging, self-reference, rupture-and-repair, emanation. Across the evening the marriage is not one feature but a family: the feature that switched on for his mother and daughters goes silent for eighty-one turns and returns as the Wife of Heaven; the feature of the marriage to uncertainty returns after seventy-eight; one feature switches on thirteen times on The Wife of Heaven does not …; two are on for most of the evening. Every time one of these comes back the sentence it peaks on is printed. Each character is then read through its own adapter, where the marriage is a single unreturning feature in him and part of the ground in her, and the paper closes with what the instrument does not see. Every feature is named twice, Neuronpedia beside ICRA; 141 features were named for this paper and the dictionary grew from 1,080 to 1,221. Full HTML text: https://icra.tanazur.org/papers/two-voice-worms/ — ICRA pre-print series: https://icra.tanazur.org/