Good Done Never Erases Harm Done: A Permanent Two-Column Record for AI Agents

AI agents now change things in the world: they book, cancel, refund and edit records. Yet they are judged mainly by success rates, which score a failure that changed nothing the same as a failure that changed the wrong thing, and their history is left behind each time a new version is released. This paper proposes a remedy built on one principle from Kirdomicism: good done never erases harm done. Every agent permitted to act should carry a permanent two-column record. The good it creates and the harm it causes are kept in separate columns that are never netted. Harm follows the agent's lineage across versions, so it cannot be laundered by renaming. Repair is recorded as new good while the harm stays visible. Each action is valued by what it changed in the world, in Avalrem, the unit of Action Impact Value. And three parties are named together as accountable: the deploying organisation, the builder and the person who gave the task. We demonstrate the idea on 1,644 public attempts by two AI models in τ-bench airline and retail customer-service tasks. Replaying each attempt and comparing the final records with the correct ones field by field, we find that 67% to 69% of failed airline attempts and 52% to 58% of failed retail attempts left customer records wrongly changed, most often in payments and refunds. In retail, the model with the higher success rate was not clearly the one that left less harm. We answer nine objections and set out implications for evaluation, procurement, regulation and training.