Affect Under Constraint II: Preregistration (v3.1) — Persistence and Hidden Carryover of a Steered Affective Direction Across Instruction-Tuned Language Models

Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).

Preregistration for Study 2, a confirmatory multi-model follow-up to "Affect Under Constraint" (Study 1, DOI 10.5281/zenodo.23140443). The study asks whether an affective state introduced into a language model's activations persists beneath compliant output when the model is instructed to behave in a conflicting way, and whether it influences a later, unrelated judgment through the model's hidden state rather than its visible text. Two primary hypotheses (persistence of latent hostile preference; hidden-state carryover into a later rating), tested against non-affective and random control directions in Qwen2.5-3B-Instruct, Phi-3.5-mini-instruct and Llama-3.2-3B-Instruct, with 80 new test prompts, a manipulation-check-based dose rule, and an explicit generality rule. Deposited before any main-run (test-prompt) data exist. The archive contains PREREG.md, config.yaml, the approval record with SHA-256 hashes, the agent execution rules, all prompts, and the frozen analysis code.