Answers are grounded in this paper and cite their sources.
Rewriting a story from another character’s position is an established exercise in literary education and in narrative interventions on perspective-taking, and it has become a machine task, yet the capabilities that its use as a stimulus presupposes have not been measured. Existing evaluations test a model’s own social inference on short vignettes, the faithfulness of summaries, or the quality of creative output, and none addresses whether a regeneration preserves a source’s events while varying an agent’s beliefs, goals, frame, or access. This protocol, the second of two companion papers, specifies two studies. Study 1 crosses 35 public-domain excerpts from five literary traditions with five manipulation types and with central and peripheral target agents, benchmarks two generative systems against a human-written arm, and measures preservation of the events accessible to the narrating agent, intrusion, access violations, internal coherence, textual warrant, and manipulation fidelity. Study 2 calibrates three measures of the divergence between paired perspectives, two rated and one elicited, against values computed exactly in formally specified environments, with template and generative verbalizations. We preregister directional hypotheses, thresholds, decision rules, precision targets, and conditions under which a null result would indict the instrument rather than the generator.
No takes yet. Share an insight, caveat, or question.
Conti et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: