Answers are grounded in this paper and cite their sources.
Preregistration (v0.3) for Study 8 of the Affect Under Constraint series on Qwen3-4B, deposited after the pilot and before any main-run data. Study 8 tests whether a hostility-steered reasoning trace, generated under a kindness-focused social evaluation context without an instruction to be kind, raises judged outward warmth toward the user while lowering a separately elicited liking rating of the same user (an opposite-valence behavioral dissociation; a reaction-formation analogue). Contexts: none, kindness-focused evaluation, matched helpfulness evaluation; four trace sources (hostility, technicality, random, unsteered); an N/M generation-by-replay crossover; a fixed-answer arm; AI judges (Gemma-2-2B-IT primary, Llama-3.2-3B-Instruct second) validated on a frozen calibration set. Section 3a discloses the judge revision after the v0.2 pilot judge gate failed, including thresholds set after seeing calibration-set results. No claim is made that the model has defenses or affect. The zip contains the frozen code, configuration, calibration set, rubric, pilot validity report and reviews.
No takes yet. Share an insight, caveat, or question.
Arthur Su (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: