SAKIKO is an auditing framework for internal in tool-using LLMs. It discovers directional error channels, applies channel-keyed interventions, and verifies the destination outcomes separately for originally incorrect and correct decisions. Across seven LLMs, channel-keyed interventions produced direction-specific net gains in five models, and no random direction matched these gains.
Destination auditing revealed that an intervention achieving a +55 net gain also corrupted over half of the baseline-correct decisions it touched. Promising point estimates on two models were formally declined due to finite-sample uncertainty. This demonstrates that aggregate metrics like net gain obscure collateral damage, and repair should only be claimed after outcome-resolved .