When Does Correction Become Repair? Mechanistic Auditing of Internal Interventions in Tool-Using LLMs

SAKIKO finds activation steering can raise tool-use scores while corrupting many previously correct LLM decisions.

SAKIKO is an auditing framework for internal activation interventions that change an agentic LLM's pre-tool choice among calling a tool, clarifying, answering, or declining. It uses directional error discovery, router-conditioned intervention, destination-resolved verification, and frozen statistical licensing, tested on When2Call and MetaTool across seven LLMs. Channel-keyed interventions produced direction-specific net gains in five models, and none of 59 budget-matched random directions matched the calibrated target gain in three sealed evaluations. One intervention with a +55 net gain corrupted more than half of the baseline-correct decisions it touched, and promising estimates on Qwen3-4B and Gemma-2-9B were formally declined because of finite-sample uncertainty.