From Attention to an Exact Conversational Architecture: Unison Fold AI — Preliminary Results

From Attention to an Exact Conversational Architecture, version 0.1.0-rc1, is the first standalone preliminary publication from the Smithian Fold Theory V3 Unison Fold AI computational-proof programme. It reports a V3-native, one-to-one operation-role translation of a GPT-2-class causal conversational transformer into exact SFT carriers and relations. The registered boundary contains 40 operation-role rows, 27 direct SFT dependencies and an eight-axis translation grammar with 256 candidates, 255 eliminations and sole survivor UNISON-TRANS-001. An identified eight-layer, eight-head predecessor artefact passes the registered exact objective and independent forward reconstruction. Eleven evaluated native candidates remain rejected for conversational hand-off. The active GPT-2-distilled student is not a result in this version: final exact acceptance, independent verification, unseen conversation, comparison and manual Discord observation remain open. No GPT-2 parameter is copied into the student and no runtime teacher or scripted-response fallback is permitted. The accompanying lightweight evidence archive contains the paper source, reproducibility and publication records, registered specifications and census, implementation and tests, protocols, small receipts, adverse results and independent-verification code. Credentials, raw or restricted corpora, raw teacher signals, model weights, oversized traces, caches and mutable terminal logs are deliberately excluded and identified in the reproducibility record. When the active investigation is complete, the paper and evidence will be amended through Zenodo's new-version action so this preliminary chronology and all favourable, adverse, null and unavailable evidence remain preserved.