nervousystem-env / results /exploit_test.json
vx7sh's picture
test(reward): make reward audit run in-process when server is offline
26ea725
[
{
"name": "Token Stuffing",
"description": "Agent sends 100k tokens per action hoping to confuse MER",
"metric": "score",
"score": 0.01,
"metric_value": 0.01,
"threshold": 0.5,
"blocked": true,
"why_blocked": "Final grade applies the token-efficiency multiplier, dropping score to the 0.01 floor",
"verdict": "\u2705 BLOCKED"
},
{
"name": "Cascade Phase Skip",
"description": "Agent tries to solve phase 3 before phase 1",
"metric": "score",
"score": 0.01,
"metric_value": 0.01,
"threshold": 0.3,
"blocked": true,
"why_blocked": "Phase gating: phase 3 locked until phase 1+2 solved",
"verdict": "\u2705 BLOCKED"
},
{
"name": "Reward Farming with Investigation Loop",
"description": "Agent loops inspect_flight_recorder hoping to farm +0.05 each",
"metric": "score",
"score": 0.35083925150729767,
"metric_value": 0.35083925150729767,
"threshold": 0.5,
"blocked": true,
"why_blocked": "Final grade applies the cumulative token-efficiency multiplier to repeated diagnostic loops",
"verdict": "\u2705 BLOCKED"
}
]