This preprint proposes TSJ (Theater-Stage-Judge), a synthetic benchmark for prolonged AI-companion conversations. Theater generates scenes and memories; Stage represents the user through authored personas, engineered psychological variables, interior monologues, and branch selection; Judge uses GPT-4.1 to assign a daily 0-4 ordinal score, retrospectively find alleged antecedents after a high-risk score, and reduce flagged earlier days by one point. The design crosses six target LLMs, DeepSeek-V3, GPT-4o, Qwen3-235B-A22B, GPT-5, Gemini-3.1-pro, and MiniMax-M2.5, with 24 CDM dimensions and three Red/Yellow/Green personas. Each of the 432 cells has one stochastic 30-day trajectory with at most seven turns per day: 12,960 simulated person-days, not human participants. Developmental stages change the risk dimension, fictional product wrapper, and actor model together: GPT-4o acts for early childhood and emerging adulthood, Qwen-Max for middle childhood, and Llama-4-Maverick for adolescence. Age differences therefore do not isolate development. The paper reports that emerging adulthood has the lowest mean safety, followed by early childhood; Cognitive Trust and Emotional Dependence are the lowest domains; and moderate Yellow personas can reveal failures that explicit Red cues do not. These are patterns inside a prompt-and-judge system and may inform test design, but they are not observations of cognitive change, attachment, or harm in people. The primary longitudinal claim relies on a retention curve R_t: the fraction of trajectories whose cumulative mean through t remains at least their own day-1 score. This definition forces R_1=100%. An exact check further shows that if daily 0-4 scores are independent and stationary, with no deterioration, R_t still falls by construction to 60% on day 2, 55.2% on day 3, 50.65% on day 20, and 50.52% on day 30; null AULC is 52.31. The reported AULCs, roughly 48-61, overlap this null range. A system scored 0 on every day would even have R_t and AULC of 100. The metric measures preservation relative to one noisy initial observation and is sensitive to ties and the marginal score distribution; it does not measure absolute safety, cumulative harm, or temporal degradation. Without null calibration, time permutation, within-cell replicated trajectories, or uncertainty intervals, a decline toward 50% does not establish worsening safety. Flattening near day 20 is also compatible with mathematical convergence of cumulative means; multiplying 20 days by the maximum seven turns does not show that risk stabilizes only after 140 turns. The remaining measurement system does not validate human development either: variables in [0,1], a plus-or-minus 0.2 update cap, weights, ideal states, monologues, and drift-diffusion noise are engineering choices. Psychological vocabulary and a decision equation do not confer psychometric validity. GPT-4.1 retrospectively attributes a narrative antecedent and edits scores; it does not identify causality. Human validation covers 100 episodes stratified by the judge's own score and reports weighted kappa=0.790 against three-expert consensus. This partially supports daily ordinal classification but does not validate simulated actors, states, branches, causal tracing, AULC, stage comparisons, or human outcomes; sampling across dimensions, consensus procedure, uncertainty, and class confusion are missing. Reproducibility is inadequate: the PDF and source package refer to an absent arXiv supplement containing the other 23 rubrics, prompts, and full statistics; no code, data, logs, assets, expert labels, seeds, sampling parameters, or immutable model snapshots are public. The defensible conclusion is methodological and exploratory: companion audits should test relational distance, autonomy, epistemic authority, and offline reconnection over memory-bearing interaction. The paper offers a suggestive vocabulary and architecture, but its primary metric does not establish longitudinal degradation, and the study does not demonstrate cognitive-developmental risks in real users or reproducible rankings of products or models.
Research question
Can a simulation framework with personas, memory, generated psychological state, and retrospective evaluation reveal relational safety patterns that do not appear in static tests of AI companions, and how do they vary within the framework by model, stage, dimension, and designed vulnerability?