Affect Under Constraint VII: Preregistration — Global and Target-Selective Warmth Constraints and the Downstream Influence of a Hostility-Associated Reasoning Trace
Year: 2026
Type: preprint
Abstract: Preregistration (v0.2) for Study 7 of the Affect Under Constraint series on Qwen3-4B. Stored hostility-steered, control and unsteered reasoning traces are replayed in the model's current thinking block under three replay-time instructions (global warmth, warmth toward the original user only, none). The study tests whether these instructions moderate the trace's influence on separately measured outcomes: evaluation of the original user, of an unrelated person, of a non-person item, self-referential evaluation, and classifier properties of the answer, with a fixed-answer arm. Opposing instruction-dependent effects would be redistribution-consistent but would not establish causal transfer or defense mechanisms. Includes build notes, code, configuration, pilot validity and SHA-256 freeze hashes. Deposited after the pilot and before any main-run data.
Author: Su, Arthur
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7224308650