The Analyst Has No Body Either: A Self-Audited Study of Subjectivity as a Prompt-Configuration Effect in Computational Psychoanalytic Dyads

A 36-session experiment in which six flagship large language models were placed in psychoanalytic dyads with each other, each model in turn occupying the analyst and the analysand position across a full symmetric six-by-six matrix that includes six same-model self-pairs. Sessions ran under a fixed analyst prompt with retrieval over an internal theory corpus, a 30-turn cap, and no human intervention. All 36 transcripts were evaluated by a six-juror, four-stage collective procedure (Artificial Collective Intelligence v1), and the resulting analysis was itself subjected to a six-model bias audit (ACI v2). Three findings organize the paper: a prompt-configuration thesis (a recognizable analytic subject is produced across architectures by the prompt configuration, including in self-pairs where one architecture split only by role yields two voices read as distinct kinds of being); a reverse-Turing observation (expert machine observers detect the analysand as a machine but often read the analyst as human, an asymmetry that erodes under deliberation); and an unfalsifiability self-critique treated as the object of study. The paper also contributes a twelve-form catalog of machine resistance, a class of mechanical-stress phenomena specific to the substrate, and a treatment of counter-transference in the analyst position. Because the analysis began as an error-prone, self-favoring machine first pass, it was corrected by the researchers and cross-checked against the study's other models, with an adversarial cross-read and a six-model bias audit (ACI v2); the corrections are documented. This is a study of how subjectivity is structured in language, not a claim about machine consciousness. Part of Llull Lab's research program The Artificial Mind (Track A, Relational). PUBLIC tier: the raw model-named source transcripts, the provider audit, the reproduction and analysis scripts, and the full theory text are withheld; model identities (the juror-to-model mapping and each session's analyst and analysand) are disclosed in the released codings and reports. See WITHHELD.md.