Evidence · DiggingBeagle record
CAGE 4 audit: discarded Restore actions charged, hidden busy state changes rankings and defender sampling causes 88.9% variance
Original CAGE 4 measurement audit reproduces defender rewards and analyzes actions dropped by busy-state handling, Restore penalties, wrapper visibility and incomplete random seeding. Reported ranking changes reflect a simulator and policy-evaluation stack, not a live network incident.
- Source role
- primary disclosure
Evidence record
Original CAGE 4 measurement audit reproduces defender rewards and analyzes actions dropped by busy-state handling, Restore penalties, wrapper visibility and incomplete random seeding. Reported ranking changes reflect a simulator and policy-evaluation stack, not a live network incident.
Claim-level citations (1)
- supportsCAGE 4 benchmark scoring can penalize agent actions the simulator never executed: The CAGE 4 audit reports 88.9% of KEEP score variance under the stock attacker came from within-episode policy-action sampling, while ignored Restore attempts could still incur penalties.
Section 4 Silent Drop; Section 5 variance decomposition and Figure 2