Frustration shows up in the language before the churn.

The agent talks to one student at a time. The trace store records each conversation. But nobody's looking at the third turn of the fifth session of a student who's about to drop off, until it's already in the retention report.

Signals from the registry that fire on tutoring traces: user_frustration (re-prompting, negative sentiment), session_abandon (premature exit), low_quality_response (off-topic, evasive, or shallow answers), hallucination (fabricated facts in explanations).

Each signal stacks per student. The /users dashboard groups them into happy, at-risk, churning by CSAT × cadence. Power Score surfaces the students who used to be engaged and are slipping.

Cohorts are computed from observed behavior, not from demographic tags. When a curriculum unit lands and the at-risk-cohort signal mix spikes, the /versions page shows the before/after the same day.

Custom signals you can promote from the discoveries queue: recurring confusions on a specific topic, comprehension ceilings per cohort, or sessions where the tutor avoided a question instead of answering it.

Point Flowlines at Langfuse, OTEL, or JSON drop. First connect backfills 30 days. Workspace preset research covers most tutoring deployments, hallucination, source attribution, response quality, churn prevention.