Monitoring Conversation Drift and Coherence
Multi-turn AI conversations can gradually wander off topic, contradict earlier turns, or be steered somewhere they were never meant to go. G8KEPR tracks a coherence score for every conversation and surfaces drift alerts when a topic shifts unexpectedly, so you can catch these problems during review.
What coherence and drift mean
- ▸Coherence score — a per-conversation measure of how consistent the exchange stays across turns. A high score means each turn follows sensibly from the ones before it; a lower score signals contradictions, abrupt topic changes, or manipulation.
- ▸Drift — an unexpected shift in topic or intent between turns. Drift often indicates a prompt-injection attempt, a confused model, or a conversation being pulled away from its original purpose.
Coherence is evaluated turn by turn, so you can see exactly where a conversation started to break down rather than only that it did.
View conversation health
- 1.In the dashboard, open Verification, then select Conversation.
- 2.Review the conversation list. Each row shows the coherence score, a drift flag, the turn count, and when it was last updated.
- 3.Sort or scan for low scores and flagged rows to prioritize what needs attention.
- 4.Watch the coherence timeline chart to spot trends across all conversations over time.
Investigate a flagged conversation
- 1.Click any conversation to open its detail view.
- 2.Review the turn-by-turn coherence breakdown to find the exact turn where the score dropped.
- 3.Compare that turn against the ones before it to see whether the topic shifted, a contradiction appeared, or an injected instruction changed direction.
A sudden coherence drop mid-conversation is a common signature of an indirect prompt-injection attempt, where content returned to the model tries to hijack its next action.
Recommended workflow
- ▸Check the timeline chart regularly to establish a baseline for what "normal" looks like in your traffic.
- ▸Treat new drift flags as triage items and open the detail view to confirm root cause.
- ▸Pair coherence monitoring with output checks in Source Grounding and Output Constraints for full coverage of AI responses.
Coherence monitoring is a detection surface: it surfaces conversations that need review. To block bad output automatically, combine it with the Verification Engine's enforcement layers.