This paper perfectly articulates the problem I spent the last year solving. The shift from "hallucination" to "fidelity decay" is the correct mental model for agent stability.
I built an open source framework called SAFi that implements the "Fidelity Meter" concept mentioned in section 4. It treats the LLM as a stochastic component in a control loop. It calculates a rolling "Alignment State" (using an Exponential Moving Average) and measures "Drift" as the vector distance from that state.
The paper discusses "Ground Erosion" where the model loses its hierarchy of values. In my system, the "Spirit" module detects this erosion and injects negative feedback to steer the agent back to the baseline. I recently red-teamed this against 845 adversarial attacks and it maintained fidelity 99.6% of the time.
It is cool to see the theoretical framework catching up to what is necessary in engineering practice.
I built an open source framework called SAFi that implements the "Fidelity Meter" concept mentioned in section 4. It treats the LLM as a stochastic component in a control loop. It calculates a rolling "Alignment State" (using an Exponential Moving Average) and measures "Drift" as the vector distance from that state.
The paper discusses "Ground Erosion" where the model loses its hierarchy of values. In my system, the "Spirit" module detects this erosion and injects negative feedback to steer the agent back to the baseline. I recently red-teamed this against 845 adversarial attacks and it maintained fidelity 99.6% of the time.
It is cool to see the theoretical framework catching up to what is necessary in engineering practice.
Repo link: https://github.com/jnamaya/SAFi