Affect Under Constraint IV: Preregistration — Does a Model Carry a Steered State Forward Because It Wrote the Text, or Because It Read It?

This record contains the preregistration for Study 4 of the Affect Under Constraint series, deposited before any main-run (test-prompt) data were collected. Study 3 found that a hostility state steered into Qwen3-4B's reasoning trace carried forward to a later judgment through the trace text. Study 4 asks whether that carryover depends on the trace being attributed to the model itself or to another assistant. Each steered trace is generated once and the identical tokens are replayed in three framings: native reasoning, quoted notes labelled as the model's own, and quoted notes labelled as another assistant's. It further tests whether the carried affect is redirected from the original user toward the attributed writer, using independent rating branches. The record includes the preregistration (v0.6), framings, configuration, approved test prompts, runner and frozen analysis code, synthetic validation, pilot validity and referent re-pilot summaries (pilot data are descriptive only), the deviations log, and the approval record with SHA-256 checksums of every binding file (approved 2026-10-24). Continues the Study 3 preregistration (10.5281/zenodo.23247883).