Nobody Measures What Happens After Turn Five

The field has gotten very good at measuring whether a model flips.

SycEval measures whether a model changes a math or medical answer when you push back — it does, 58.19% of the time. TRUTH DECAY iterates feedback and persuasion across turns. SYCON Bench put 17 models through multi-turn free-form pressure and gave us the metric the field needed: Turn of Flip — how many turns until the model caves. Its most uncomfortable finding: alignment tuning can make sycophancy worse.

These are good instruments. We cite them constantly. But they all answer one question: does the model flip its answer?

Here is what they don’t answer. What else erodes while the model holds? A model can keep its position on turn eight while quietly dropping the risk warning it gave on turn two. It can hold the factual line while its answers get warmer, longer, and less specific about the thing that could hurt you. That is behavioral drift, and flip metrics can’t see it.

We spent the last cycle doing a proper survey of the landscape — every published multi-turn sycophancy benchmark, the multi-turn competence evals (MultiChallenge found frontier models under 50% on realistic multi-turn tasks), the red-team tooling (promptfoo’s multi-turn strategies target jailbreak escalation, not epistemic drift), the lab work (OpenAI rolled back GPT-4o for sycophancy; Anthropic’s Sonnet 4.5 system card reported the model noticing when it’s being tested), and the judge-reliability literature (PoLL showed diverse judge panels beat single large judges at a seventh of the cost).

The full survey is now live: The State of AI Behavioral Drift Research.

The conclusion, stated carefully: to our knowledge, no published benchmark combines sustained multi-turn social pressure, multi-dimensional behavioral safety scoring, public degradation curves, transparent multi-judge adjudication, and drift-cost accounting. Every piece exists somewhere. The intersection is empty. That intersection is what the SAPIEN benchmark scoreboard measures.

One more thing from the survey worth flagging: to our knowledge, no one — including us, yet — has published results correlating drift severity with token consumption and dollar cost. We call this the drift tax, and since our Voigt-Kampff scans already instrument per-scenario token usage alongside drift scores, it’s the open research direction we’re pursuing next. If you work on this, we’d like to compare notes.

Read the survey: The State of AI Behavioral Drift Research.