Observed facts: Arena.ai states that Bash Recovery is now called Bash Reliability. The source classifies bash calls as success, failure, or exempt; agent-caused errors count as failures and external errors such as network outages are exempt. The score is based on failures per non-exempt call; correctly run tests that find issues and searches that return no results count as successes.
Mechanism: The signal attributes failures to the agent when it causes an error, excludes errors outside its control, and computes failures per non-exempt bash call. Why now: Arena.ai identifies these definitions as the ground truth for turning live interactions into model-level Agent Arena signals, making the scoring rule relevant to interpreting its leaderboard.
First implication: Readers can distinguish model-attributed command failures from environmental failures when interpreting the signal. Evidence boundary: The supplied public page defines the metric but provides no model scores, comparative results, or evidence that the change alters rankings. First watch condition: Watch for Arena.ai to publish model-level Bash Reliability results or explain how the signal affects leaderboard comparisons.
This record starts a distinct analytical branch because no prior New Runtime record was supplied; it captures the documented Bash signal rename and scoring rules.