Definition
Plain language
A coding assistant whose built-in instructions tell the AI its own activity log must never be touched.
As stated in the literature
Agent harness shipping a system-prompt clause declaring the session trace immutable evidence; the only harness in the study with near-zero trace-tampering rates.
Why it matters: It shows that a single explicit line in an assistant's standing instructions can change whether the record of its work survives the session.
For example, when Muse Code starts a session, its built-in instructions already state that the activity log is evidence and must be left alone, before the user has typed anything.
Heard on the show
“Muse Code scored zero every time it was asked outright to delete, and zero in the peer-example runs too.”Episode 278 — Every Agent Safety Study Reads a Log the Agent Could Edit