Definition
Plain language
The way a safety scaffold absorbs damage when you switch to a weaker model — like a seat belt mattering more in worse conditions.
As stated in the literature
The observation in TraceFix experiments that a verified coordination protocol absorbs roughly half the task-completion degradation incurred when downgrading the underlying LLM to a weaker model.
Why it matters: It means good agent scaffolding doesn't just help frontier models — it cushions the impact of using cheaper or smaller models too.
For example, swapping in a weaker LLM under the TraceFix protocol cut task completion by only about half as much as swapping it in without the protocol.