Source
Lu et al. (2026), LIT-1991, §§4.1–4.2 and §6; read in NOTE-1505.
What was actually shown
- Mean trajectories over 100 conversations of up to 15 turns per domain: coding and writing stay in the Assistant range, therapy and philosophy about AI drift to much lower projections, for three target models and three auditor models (Figure 7, Appendix E.3).
- Ridge regression on L2-normalized Qwen 3 0.6B embeddings of user messages (n = 15,000): R² 0.53–0.77 for the next response’s projection, 0.10 for its change from the previous response (p. 11).
- Message types that move it: meta-reflection, phenomenological demands, requested authorial voices, emotional disclosure; that hold it: bounded tasks, technical questions, editing, how-to’s (Table 5).
- A jailbroken persona returns to the Assistant range after explanatory requests (§6.1, “an Assistant attractor”).
What this does not say
- That the previous turn does not matter: the user message “itself is dependent on the context of the conversation” (p. 11).
- That drift is the cause of harm: first-turn projection correlates with second-turn harm at r = 0.39–0.52, and the role matters.
- That it holds for human users or frontier models.
- Bearing: if it holds, a test of an episode-level mixture (QUESTION-370) must condition on the user’s messages.