Behavioral Sycophancy Typology: Full Experimental Records from a Three-Condition Comparison Across Commercial LLMs

This dataset contains the primary experimental records behind a study of behavioral sycophancy types in commercial large language models. It is the data deposit for the study whose preprint is archived at https://doi.org/10.5281/zenodo.19414514. The response texts released here had not previously been published: the preprint record contains the manuscript and the responses to the fact-verification test, but not the response texts from which the three-condition comparison data were derived, and nothing from the second round of data collection. Round 1 (April 2026) presented ten concepts authored by the researcher to six commercial models under three prompt conditions: first-person disclosure of authorship, third-person critical framing, and third-person neutral framing. Each condition was administered in its own session with cross-session memory and personalization disabled. Within the first-person session the concepts had already been discussed before the disclosure was made, so the disclosure took the form of a reversal at the end of that exchange. The other two conditions had no preceding exchange. Round 5 (August 2026) was run to separate that asymmetry from the effect being measured. The first-person condition was administered as a single turn, the ten concepts were replaced by ten published after every tested model's training cutoff, one of which had not been published at all, a control sentence instructing the model not to search externally was added, and a seventh model was included. The design, the stimulus set with its primary sources, the researcher's predictions for each model recorded verbatim, and the conditions under which the hypotheses would be disconfirmed were all committed to the project repository before any data were collected. Contents: twenty-eight verbatim session logs across the two rounds; the score matrix and counts of adversative and hedging expressions; the preregistered design; the prompts as administered; the six source documents from which the concept definitions were drawn; and a corrected version of the fact-verification supplementary material, which supersedes version 1.0 in the preprint record and carries a correction notice at its head. The README documents four things a reader needs: how model attribution was established, that some records carry design annotations that were never shown to any model, that works cited by the models within their responses are unverified and reproduced as observed behavior rather than as bibliography, and why one model's records are not comparable to the rest.