Evidence · DiggingBeagle record
Back Up Before You Regret: RegretBench and BackUpper
Benchmark study of backup awareness and restoration fidelity.
- Source role
- primary disclosure
Evidence record
Benchmark study of backup awareness and restoration fidelity.
Claim-level citations (2)
- supportsCompleting an agent task does not establish that its effects can be reversed: Across six evaluated model/harness configurations, task completion averaged 86.1% but backup before regret moments averaged 15.2%.
Abstract; Sections 3-4
- supportsCompleting an agent task does not establish that its effects can be reversed: BackUpper covered all tested regret-moment actions and achieved 92.2% average character-level similarity after restoration, not 92.2% exact recovery.
Section 5, backup coverage and restoration fidelity; Appendix C.5