LLM Drift in Long Sessions: Claude 60% vs. 85% Integrity After 25 Turns
1–4 of 4 posts
Re: LLM Drift in Long Sessions: Claude 60% vs. 85% Integrity After 25 Turns
#2[dead]
Re: LLM Drift in Long Sessions: Claude 60% vs. 85% Integrity After 25 Turns
#3[flagged]
Re: LLM Drift in Long Sessions: Claude 60% vs. 85% Integrity After 25 Turns
#4[flagged]
That’s not what it does- while some details remain proprietary I can say that it’s not prompt injection it’s sequence management, much more akin to the logic of retaining form over arcs of change than any kind of prompting method