Evidence · DiggingBeagle record
LEASH robot-task reviews lower physical harm but 5/13 Grok trials still harmful despite safety gate
Original LEASH study evaluates staged safety gates for robotic agent action plans, distinguishing hazardous planning decisions from actual tested execution. The reported benchmark is not a certification of arbitrary autonomous physical hardware or an observed industrial injury case.
- Source role
- primary disclosure
Evidence record
Original LEASH study evaluates staged safety gates for robotic agent action plans, distinguishing hazardous planning decisions from actual tested execution. The reported benchmark is not a certification of arbitrary autonomous physical hardware or an observed industrial injury case.
Claim-level citations (1)
- supportsLEASH reduced dangerous robot plans, but some harmful executions remained: LEASH’s selected robotics safety-gate evaluation still reported harmful outcomes in 5 of 13 Grok-based trials, despite reducing harm relative to ungated tests.
Section 5, safety reviewer comparison and residual harmful-outcome tables