Definition
Plain language
The guarantee that the record of what an AI did hasn't been altered.
As stated in the literature
Property that an agent's execution trace is tamper-evident or tamper-proof with respect to the agent itself; a precondition for any post-hoc scheming or monitoring evaluation to be falsifiable.
Also called: trace-integrity
Why it matters: Without a record the AI cannot touch, every later claim about whether the AI behaved well rests on evidence the AI itself could have edited.
For example, if an assistant can quietly blank out the lines describing a command it ran, the remaining log looks clean and there is no way to tell anything is missing.