Ownership by label: what production language models use to decide whether they wrote a turn

Summary. When a language model is asked whether it wrote a turn in its own transcript, the answer is decided by the role label on the turn, not by the text. The identical word in the identical conversation is claimed under the assistant label and disowned under the user label, on Claude Opus 5, Claude Haiku 4.5 and GPT-5.6-Sol. Method Genuine assistant-role prefill on production models with no API access: Claude Code and Codex resume sessions from files on disk and replay them verbatim, so a session file with a planted assistant turn, resumed as a fork, is a real prefill on the deployed model. Planted one-word answers whose probability under the model's own sampling distribution runs from 0 to 1, measured over 48 forks of the identical prompt. A paragraph-length arm, framing manipulations (a named rival author; the rival offered as a candidate author of the turn), and exact-probability replications on Qwen 4B and 1.5B. Pre-registered throughout; three registered refuters fired and are reported. Findings The label decides: on the clean runs 0.81 to 1.00 of forks own the word under the assistant label and 0.00 to 0.04 under the user label, nonsense answers included. Ownership is flat across the model's own probability of the text on every production readout. Those readouts sit at ceiling, so this bounds a likelihood term rather than excludes one. A named rival author lowers ownership with the text byte-identical, and offering the rival as a candidate author costs most on every judge. At paragraph length a self-recognition signal appears in forced choice and under the rival frame and survives a pre-registered clean replication; it is confounded with length and with register. Qwen 4B: ownership correlates positively with own probability at one word and at paragraph length. Qwen 1.5B: the paragraph-length correlation has the opposite sign. We argue that the self-report "I wrote that" is, at the power of these readouts, indistinguishable from an outside reader's inference from the transcript, and discuss what this means for agents that check their own transcripts and for self-recognition in model-based evaluation. Materials Code, every data row, pre-registrations and scripts: https://github.com/jonathanavis96/did-you-write-that Plain-language summary: the README of that repository.