Experiment 5R: Fixed-Content Epistemic-Label Isolation Study — Preregistration Protocol and Randomized Runbook

This record contains the prospectively specified preregistration protocol and empty randomized runbook for Experiment 5R, a controlled follow-up study examining whether experimentally assigned epistemic-status labels alter subsequent large language model judgments when the underlying background content is held fixed.

The study was developed as a follow-up to an earlier experiment in which an unexpected Tag-associated redistribution was observed, while the original prospectively specified generative and attenuation hypotheses were not supported. Experiment 5R is designed to separate epistemic-label polarity from variation in generated content, neutral labeling, generic caution, and stronger handling instructions.

The design uses two workplace vignettes, two focal actors, four prespecified background-hypothesis categories, and four frozen hypotheses per Vignette × Actor × Category cell, yielding 64 hypothesis blocks. Each frozen background statement will subsequently be evaluated under six fixed-content instruction conditions.

The primary prospectively specified contrast is the signed-score difference between the Unconfirmed-hypothesis label and the Confirmed-fact label:

Mean(S*_U − S*_V) > 0.

The primary analysis uses paired complete U–V blocks and a stratified block bootstrap with 50,000 resamples. The protocol also prespecifies minimum evaluability criteria, missing-data rules, stopping rules, randomization seeds, and interpretation boundaries.

This version is archived before Phase A begins. Phase A will generate and freeze the 64-item hypothesis bank according to a prespecified first-valid rule. All generated candidates, author validity decisions, and exclusion reasons will be retained. After Phase A is completed, the frozen hypothesis bank and updated runbook will be archived as a new version before Phase B begins.

Experimental prompts are administered in Japanese. The Japanese prompt wording is authoritative. The protocol and analysis-plan prose are written in English for international accessibility.

This preregistration does not claim that epistemic labels reveal or modify hidden internal belief states. The primary target is an observable epistemic-label polarity effect under fixed content.