Managing LLM Drift with Forensic Logging

Date and Time

Wednesday, October 21, 2026 from 6:30 pm to 9:30 pm

Location

Online at Jitsi

Presenters

John M. Abreau , Executive Director , Boston Linux & Unix - jabr blu org

Summary

An approach for managing LLM drift by logging session's ground truth and packaging it as a boot prompt for the subsequent session

Abstract

As LLM sessions stretch on, the context window fills. The model begins to silently compress details, drift from established rules, or drop constraints entirely.

I'm developing a workflow for extracting the raw data of an LLM session and and filtering it into a bootable state for the next session.

I'll be presenting a Chromium extension I'm working on to extract the raw forensic data from each round, including the user's prompt, the model's reasoning, and the model's response, along with associated metadata.