Provenance and Authorization in Autobiographical Language Models

Autobiographical language models may draw on archives containing observations, quotations, emotional reports, and earlier model outputs. Cryptographic authentication can preserve an archive while leaving these sources semantically conflated. We formulate this gap as a computer-security problem: narrative content must not silently become consent or alter an authorized decision rule. We define typed records, authenticated attestations, a separate policy channel, and admissible views. Under explicit assumptions, we give a standard integrity reduction and a noninterference argument, characterize the information required by exact decision-preserving summaries, and show why passive history and fixed-point convergence need not identify a preferred action. Alpay Algebra supplies a comparative language for iterative memory without being a premise of the proofs. Sociology of religious affiliation, source-monitoring psychology, and self-determination theory motivate distinctions in a bounded Latter-day Saint vignette. A separate, reproducible pilot uses a frozen public language model on 32 wholly synthetic cases across 576 paired conditions. It compares flattened records, source labels, and an explicit admission gate under contradictory summary repetition. The study concerns computational attribution and authorization, not the motives or psychology of the people in the vignette.