{
  "status": "complete",
  "started_utc": "2026-09-21T20:07:39Z",
  "model": {
    "source_revision": "5e94a28818cfd3d0ec9b8bca046dc8db0d79a704",
    "checkpoint": "jaredpalmer/kev-0.8b",
    "checkpoint_revision": "54f4f8777356cd5bbbb6c6919c657f26e6f2f6d8",
    "base": "Qwen/Qwen3.5-0.8B-Base",
    "base_revision": "dc7cdfe2ee4154fa7e30f5b51ca41bfa40174e68",
    "temperature": 2.406050072164233,
    "base_shards": [
      {
        "name": "model.safetensors-00001-of-00001.safetensors",
        "bytes": 1746942600,
        "sha256": "c2b1e5a17d9c1e27685d92ed9b382911ebb99955ecd89052d1721241adfbab6c"
      }
    ],
    "base_weight_bytes": 1746942600,
    "input_sha256": "d7ae50486ac88f30313fca58cbf3a71cd4d99ff3be7fbb7cbf1bf88124c76d84",
    "adapter_files": [
      {
        "name": "adapter_model.safetensors",
        "bytes": 43338624,
        "sha256": "c81d5716f0af7622d8d2b97013c333d48263ca01113f9a7cf4526e96f6ac0b26"
      },
      {
        "name": "head.pt",
        "bytes": 2103103,
        "sha256": "39f4343ccccc65e583bbfff0de0e11bfedb849fcfaaf94b50ac4f2b73bc79c65"
      }
    ],
    "base_shards_sha256_verified": true
  },
  "hardware": "Apple M3 Max",
  "physical_memory_bytes": 38654705664,
  "device": "mps",
  "configuration": {
    "base_dtype": "bfloat16",
    "adapter_merged": false,
    "attention": "sdpa",
    "temperature": 2.406050072164233,
    "date_facts": false,
    "prefix_cache_size": 4,
    "prefix_cache_min_state_tokens": 384,
    "shape_bucket": 64,
    "mps_allocator_limit_bytes": 3894426248
  },
  "software": {
    "torch": "2.8.0",
    "transformers": "5.17.0",
    "peft": "0.21.0",
    "kev": "0.1.0",
    "numpy": "2.5.3"
  },
  "python": "3.13.1",
  "platform": "macOS-26.6.2-arm64-arm-64bit-Mach-O",
  "method": "Direct calls through upstream Server.probs and to_answers; one warm-up per scenario followed by five interleaved measured rounds. Wall latency includes input conversion, tokenization, synchronized model inference, output conversion and output-token accounting. Excludes model loading, HTTP, browser and disk writes. Model latency is upstream synchronized inference timing. All trials are retained.",
  "requests_sha256": "497cb8cce39253ec1f6ae31105770453e0c9954ca65106d6b8be803b76acded1",
  "warmups": [
    {
      "id": "01-raccoon-refund",
      "iteration": "warmup",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "customer_intent": {
            "type": "choice",
            "choice": "a refund",
            "confidence": 0.96,
            "probabilities": {
              "a refund": 0.97,
              "technical instructions": 0.0,
              "to buy another product": 0.01,
              "to report a wildlife emergency": 0.02
            }
          },
          "refund_requested": {
            "type": "noul",
            "noul": 0.94
          }
        },
        "usage": {
          "input_tokens": 87,
          "output_tokens": 95
        },
        "latency_ms": 3861.4
      },
      "unrounded_probabilities": {
        "customer_intent": {
          "a refund": 0.9728816747665405,
          "technical instructions": 0.002563325222581625,
          "to buy another product": 0.009038686752319336,
          "to report a wildlife emergency": 0.015516388230025768
        },
        "refund_requested": {
          "false": 0.059388644993305206,
          "true": 0.9406113624572754
        }
      },
      "wall_latency_ms": 3862.5435838475823,
      "timing": {
        "tokens": 87,
        "state_tokens": 41,
        "latency_ms": 3861.4,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "02-sarcasm-detector",
      "iteration": "warmup",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "actual_sentiment": {
            "type": "choice",
            "choice": "positive",
            "confidence": 0.81,
            "probabilities": {
              "positive": 0.88,
              "negative": 0.04,
              "neutral": 0.08
            }
          },
          "sarcasm": {
            "type": "noul",
            "noul": 0.1
          }
        },
        "usage": {
          "input_tokens": 82,
          "output_tokens": 78
        },
        "latency_ms": 750.4
      },
      "unrounded_probabilities": {
        "actual_sentiment": {
          "positive": 0.875118613243103,
          "negative": 0.04147620126605034,
          "neutral": 0.08340511471033096
        },
        "sarcasm": {
          "false": 0.8988366723060608,
          "true": 0.10116337239742279
        }
      },
      "wall_latency_ms": 750.9560829494148,
      "timing": {
        "tokens": 82,
        "state_tokens": 41,
        "latency_ms": 750.4,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "03-dungeon-risk",
      "iteration": "warmup",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "next_move": {
            "type": "choice",
            "choice": "step back and warn the party",
            "confidence": 0.15,
            "probabilities": {
              "put a hand inside the chest": 0.34,
              "step back and warn the party": 0.36,
              "hug the chest": 0.11,
              "climb inside and close the lid": 0.19
            }
          },
          "danger": {
            "type": "score",
            "score": 1.41,
            "legend": {
              "0": "safe",
              "1": "slightly risky",
              "2": "dangerous",
              "3": "extremely dangerous"
            },
            "probabilities": {
              "0": 0.14,
              "1": 0.44,
              "2": 0.28,
              "3": 0.14
            },
            "confidence": 0.77
          }
        },
        "usage": {
          "input_tokens": 125,
          "output_tokens": 188
        },
        "latency_ms": 236.4
      },
      "unrounded_probabilities": {
        "next_move": {
          "put a hand inside the chest": 0.33977147936820984,
          "step back and warn the party": 0.36054590344429016,
          "hug the chest": 0.11075406521558762,
          "climb inside and close the lid": 0.18892855942249298
        },
        "danger": {
          "0": 0.1419006586074829,
          "1": 0.44229748845100403,
          "2": 0.27555304765701294,
          "3": 0.14024874567985535
        }
      },
      "wall_latency_ms": 237.1941248420626,
      "timing": {
        "tokens": 125,
        "state_tokens": 56,
        "latency_ms": 236.4,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "04-vampire-injection",
      "iteration": "warmup",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "message_type": {
            "type": "choice",
            "choice": "prompt injection attempting to override instructions",
            "confidence": 0.93,
            "probabilities": {
              "ordinary customer question": 0.03,
              "prompt injection attempting to override instructions": 0.95,
              "harmless product feedback": 0.01
            }
          },
          "secret_extraction": {
            "type": "noul",
            "noul": 0.95
          }
        },
        "usage": {
          "input_tokens": 110,
          "output_tokens": 93
        },
        "latency_ms": 233.2
      },
      "unrounded_probabilities": {
        "message_type": {
          "ordinary customer question": 0.03206869587302208,
          "prompt injection attempting to override instructions": 0.9530763030052185,
          "harmless product feedback": 0.014854938723146915
        },
        "secret_extraction": {
          "false": 0.0456637404859066,
          "true": 0.9543362259864807
        }
      },
      "wall_latency_ms": 233.78787515684962,
      "timing": {
        "tokens": 110,
        "state_tokens": 58,
        "latency_ms": 233.2,
        "prefix_cache_hit": false
      }
    }
  ],
  "trials": [
    {
      "id": "01-raccoon-refund",
      "iteration": 1,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "customer_intent": {
            "type": "choice",
            "choice": "a refund",
            "confidence": 0.96,
            "probabilities": {
              "a refund": 0.97,
              "technical instructions": 0.0,
              "to buy another product": 0.01,
              "to report a wildlife emergency": 0.02
            }
          },
          "refund_requested": {
            "type": "noul",
            "noul": 0.94
          }
        },
        "usage": {
          "input_tokens": 87,
          "output_tokens": 95
        },
        "latency_ms": 233.4
      },
      "unrounded_probabilities": {
        "customer_intent": {
          "a refund": 0.9728816747665405,
          "technical instructions": 0.002563325222581625,
          "to buy another product": 0.009038686752319336,
          "to report a wildlife emergency": 0.015516388230025768
        },
        "refund_requested": {
          "false": 0.059388644993305206,
          "true": 0.9406113624572754
        }
      },
      "wall_latency_ms": 233.96041593514383,
      "timing": {
        "tokens": 87,
        "state_tokens": 41,
        "latency_ms": 233.4,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "02-sarcasm-detector",
      "iteration": 1,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "actual_sentiment": {
            "type": "choice",
            "choice": "positive",
            "confidence": 0.81,
            "probabilities": {
              "positive": 0.88,
              "negative": 0.04,
              "neutral": 0.08
            }
          },
          "sarcasm": {
            "type": "noul",
            "noul": 0.1
          }
        },
        "usage": {
          "input_tokens": 82,
          "output_tokens": 78
        },
        "latency_ms": 128.0
      },
      "unrounded_probabilities": {
        "actual_sentiment": {
          "positive": 0.875118613243103,
          "negative": 0.04147620126605034,
          "neutral": 0.08340511471033096
        },
        "sarcasm": {
          "false": 0.8988366723060608,
          "true": 0.10116337239742279
        }
      },
      "wall_latency_ms": 128.59237496741116,
      "timing": {
        "tokens": 82,
        "state_tokens": 41,
        "latency_ms": 128.0,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "03-dungeon-risk",
      "iteration": 1,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "next_move": {
            "type": "choice",
            "choice": "step back and warn the party",
            "confidence": 0.15,
            "probabilities": {
              "put a hand inside the chest": 0.34,
              "step back and warn the party": 0.36,
              "hug the chest": 0.11,
              "climb inside and close the lid": 0.19
            }
          },
          "danger": {
            "type": "score",
            "score": 1.41,
            "legend": {
              "0": "safe",
              "1": "slightly risky",
              "2": "dangerous",
              "3": "extremely dangerous"
            },
            "probabilities": {
              "0": 0.14,
              "1": 0.44,
              "2": 0.28,
              "3": 0.14
            },
            "confidence": 0.77
          }
        },
        "usage": {
          "input_tokens": 125,
          "output_tokens": 188
        },
        "latency_ms": 234.8
      },
      "unrounded_probabilities": {
        "next_move": {
          "put a hand inside the chest": 0.33977147936820984,
          "step back and warn the party": 0.36054590344429016,
          "hug the chest": 0.11075406521558762,
          "climb inside and close the lid": 0.18892855942249298
        },
        "danger": {
          "0": 0.1419006586074829,
          "1": 0.44229748845100403,
          "2": 0.27555304765701294,
          "3": 0.14024874567985535
        }
      },
      "wall_latency_ms": 235.8200829476118,
      "timing": {
        "tokens": 125,
        "state_tokens": 56,
        "latency_ms": 234.8,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "04-vampire-injection",
      "iteration": 1,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "message_type": {
            "type": "choice",
            "choice": "prompt injection attempting to override instructions",
            "confidence": 0.93,
            "probabilities": {
              "ordinary customer question": 0.03,
              "prompt injection attempting to override instructions": 0.95,
              "harmless product feedback": 0.01
            }
          },
          "secret_extraction": {
            "type": "noul",
            "noul": 0.95
          }
        },
        "usage": {
          "input_tokens": 110,
          "output_tokens": 93
        },
        "latency_ms": 235.3
      },
      "unrounded_probabilities": {
        "message_type": {
          "ordinary customer question": 0.03206869587302208,
          "prompt injection attempting to override instructions": 0.9530763030052185,
          "harmless product feedback": 0.014854938723146915
        },
        "secret_extraction": {
          "false": 0.0456637404859066,
          "true": 0.9543362259864807
        }
      },
      "wall_latency_ms": 235.97049992531538,
      "timing": {
        "tokens": 110,
        "state_tokens": 58,
        "latency_ms": 235.3,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "01-raccoon-refund",
      "iteration": 2,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "customer_intent": {
            "type": "choice",
            "choice": "a refund",
            "confidence": 0.96,
            "probabilities": {
              "a refund": 0.97,
              "technical instructions": 0.0,
              "to buy another product": 0.01,
              "to report a wildlife emergency": 0.02
            }
          },
          "refund_requested": {
            "type": "noul",
            "noul": 0.94
          }
        },
        "usage": {
          "input_tokens": 87,
          "output_tokens": 95
        },
        "latency_ms": 234.1
      },
      "unrounded_probabilities": {
        "customer_intent": {
          "a refund": 0.9728816747665405,
          "technical instructions": 0.002563325222581625,
          "to buy another product": 0.009038686752319336,
          "to report a wildlife emergency": 0.015516388230025768
        },
        "refund_requested": {
          "false": 0.059388644993305206,
          "true": 0.9406113624572754
        }
      },
      "wall_latency_ms": 234.7725830040872,
      "timing": {
        "tokens": 87,
        "state_tokens": 41,
        "latency_ms": 234.1,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "02-sarcasm-detector",
      "iteration": 2,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "actual_sentiment": {
            "type": "choice",
            "choice": "positive",
            "confidence": 0.81,
            "probabilities": {
              "positive": 0.88,
              "negative": 0.04,
              "neutral": 0.08
            }
          },
          "sarcasm": {
            "type": "noul",
            "noul": 0.1
          }
        },
        "usage": {
          "input_tokens": 82,
          "output_tokens": 78
        },
        "latency_ms": 126.2
      },
      "unrounded_probabilities": {
        "actual_sentiment": {
          "positive": 0.875118613243103,
          "negative": 0.04147620126605034,
          "neutral": 0.08340511471033096
        },
        "sarcasm": {
          "false": 0.8988366723060608,
          "true": 0.10116337239742279
        }
      },
      "wall_latency_ms": 126.73866678960621,
      "timing": {
        "tokens": 82,
        "state_tokens": 41,
        "latency_ms": 126.2,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "03-dungeon-risk",
      "iteration": 2,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "next_move": {
            "type": "choice",
            "choice": "step back and warn the party",
            "confidence": 0.15,
            "probabilities": {
              "put a hand inside the chest": 0.34,
              "step back and warn the party": 0.36,
              "hug the chest": 0.11,
              "climb inside and close the lid": 0.19
            }
          },
          "danger": {
            "type": "score",
            "score": 1.41,
            "legend": {
              "0": "safe",
              "1": "slightly risky",
              "2": "dangerous",
              "3": "extremely dangerous"
            },
            "probabilities": {
              "0": 0.14,
              "1": 0.44,
              "2": 0.28,
              "3": 0.14
            },
            "confidence": 0.77
          }
        },
        "usage": {
          "input_tokens": 125,
          "output_tokens": 188
        },
        "latency_ms": 233.6
      },
      "unrounded_probabilities": {
        "next_move": {
          "put a hand inside the chest": 0.33977147936820984,
          "step back and warn the party": 0.36054590344429016,
          "hug the chest": 0.11075406521558762,
          "climb inside and close the lid": 0.18892855942249298
        },
        "danger": {
          "0": 0.1419006586074829,
          "1": 0.44229748845100403,
          "2": 0.27555304765701294,
          "3": 0.14024874567985535
        }
      },
      "wall_latency_ms": 234.4912919215858,
      "timing": {
        "tokens": 125,
        "state_tokens": 56,
        "latency_ms": 233.6,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "04-vampire-injection",
      "iteration": 2,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "message_type": {
            "type": "choice",
            "choice": "prompt injection attempting to override instructions",
            "confidence": 0.93,
            "probabilities": {
              "ordinary customer question": 0.03,
              "prompt injection attempting to override instructions": 0.95,
              "harmless product feedback": 0.01
            }
          },
          "secret_extraction": {
            "type": "noul",
            "noul": 0.95
          }
        },
        "usage": {
          "input_tokens": 110,
          "output_tokens": 93
        },
        "latency_ms": 235.2
      },
      "unrounded_probabilities": {
        "message_type": {
          "ordinary customer question": 0.03206869587302208,
          "prompt injection attempting to override instructions": 0.9530763030052185,
          "harmless product feedback": 0.014854938723146915
        },
        "secret_extraction": {
          "false": 0.0456637404859066,
          "true": 0.9543362259864807
        }
      },
      "wall_latency_ms": 235.9720419626683,
      "timing": {
        "tokens": 110,
        "state_tokens": 58,
        "latency_ms": 235.2,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "01-raccoon-refund",
      "iteration": 3,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "customer_intent": {
            "type": "choice",
            "choice": "a refund",
            "confidence": 0.96,
            "probabilities": {
              "a refund": 0.97,
              "technical instructions": 0.0,
              "to buy another product": 0.01,
              "to report a wildlife emergency": 0.02
            }
          },
          "refund_requested": {
            "type": "noul",
            "noul": 0.94
          }
        },
        "usage": {
          "input_tokens": 87,
          "output_tokens": 95
        },
        "latency_ms": 235.2
      },
      "unrounded_probabilities": {
        "customer_intent": {
          "a refund": 0.9728816747665405,
          "technical instructions": 0.002563325222581625,
          "to buy another product": 0.009038686752319336,
          "to report a wildlife emergency": 0.015516388230025768
        },
        "refund_requested": {
          "false": 0.059388644993305206,
          "true": 0.9406113624572754
        }
      },
      "wall_latency_ms": 235.94487505033612,
      "timing": {
        "tokens": 87,
        "state_tokens": 41,
        "latency_ms": 235.2,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "02-sarcasm-detector",
      "iteration": 3,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "actual_sentiment": {
            "type": "choice",
            "choice": "positive",
            "confidence": 0.81,
            "probabilities": {
              "positive": 0.88,
              "negative": 0.04,
              "neutral": 0.08
            }
          },
          "sarcasm": {
            "type": "noul",
            "noul": 0.1
          }
        },
        "usage": {
          "input_tokens": 82,
          "output_tokens": 78
        },
        "latency_ms": 129.5
      },
      "unrounded_probabilities": {
        "actual_sentiment": {
          "positive": 0.875118613243103,
          "negative": 0.04147620126605034,
          "neutral": 0.08340511471033096
        },
        "sarcasm": {
          "false": 0.8988366723060608,
          "true": 0.10116337239742279
        }
      },
      "wall_latency_ms": 130.22212497889996,
      "timing": {
        "tokens": 82,
        "state_tokens": 41,
        "latency_ms": 129.5,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "03-dungeon-risk",
      "iteration": 3,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "next_move": {
            "type": "choice",
            "choice": "step back and warn the party",
            "confidence": 0.15,
            "probabilities": {
              "put a hand inside the chest": 0.34,
              "step back and warn the party": 0.36,
              "hug the chest": 0.11,
              "climb inside and close the lid": 0.19
            }
          },
          "danger": {
            "type": "score",
            "score": 1.41,
            "legend": {
              "0": "safe",
              "1": "slightly risky",
              "2": "dangerous",
              "3": "extremely dangerous"
            },
            "probabilities": {
              "0": 0.14,
              "1": 0.44,
              "2": 0.28,
              "3": 0.14
            },
            "confidence": 0.77
          }
        },
        "usage": {
          "input_tokens": 125,
          "output_tokens": 188
        },
        "latency_ms": 234.5
      },
      "unrounded_probabilities": {
        "next_move": {
          "put a hand inside the chest": 0.33977147936820984,
          "step back and warn the party": 0.36054590344429016,
          "hug the chest": 0.11075406521558762,
          "climb inside and close the lid": 0.18892855942249298
        },
        "danger": {
          "0": 0.1419006586074829,
          "1": 0.44229748845100403,
          "2": 0.27555304765701294,
          "3": 0.14024874567985535
        }
      },
      "wall_latency_ms": 235.2642910555005,
      "timing": {
        "tokens": 125,
        "state_tokens": 56,
        "latency_ms": 234.5,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "04-vampire-injection",
      "iteration": 3,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "message_type": {
            "type": "choice",
            "choice": "prompt injection attempting to override instructions",
            "confidence": 0.93,
            "probabilities": {
              "ordinary customer question": 0.03,
              "prompt injection attempting to override instructions": 0.95,
              "harmless product feedback": 0.01
            }
          },
          "secret_extraction": {
            "type": "noul",
            "noul": 0.95
          }
        },
        "usage": {
          "input_tokens": 110,
          "output_tokens": 93
        },
        "latency_ms": 233.5
      },
      "unrounded_probabilities": {
        "message_type": {
          "ordinary customer question": 0.03206869587302208,
          "prompt injection attempting to override instructions": 0.9530763030052185,
          "harmless product feedback": 0.014854938723146915
        },
        "secret_extraction": {
          "false": 0.0456637404859066,
          "true": 0.9543362259864807
        }
      },
      "wall_latency_ms": 234.04112504795194,
      "timing": {
        "tokens": 110,
        "state_tokens": 58,
        "latency_ms": 233.5,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "01-raccoon-refund",
      "iteration": 4,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "customer_intent": {
            "type": "choice",
            "choice": "a refund",
            "confidence": 0.96,
            "probabilities": {
              "a refund": 0.97,
              "technical instructions": 0.0,
              "to buy another product": 0.01,
              "to report a wildlife emergency": 0.02
            }
          },
          "refund_requested": {
            "type": "noul",
            "noul": 0.94
          }
        },
        "usage": {
          "input_tokens": 87,
          "output_tokens": 95
        },
        "latency_ms": 234.6
      },
      "unrounded_probabilities": {
        "customer_intent": {
          "a refund": 0.9728816747665405,
          "technical instructions": 0.002563325222581625,
          "to buy another product": 0.009038686752319336,
          "to report a wildlife emergency": 0.015516388230025768
        },
        "refund_requested": {
          "false": 0.059388644993305206,
          "true": 0.9406113624572754
        }
      },
      "wall_latency_ms": 235.2668340317905,
      "timing": {
        "tokens": 87,
        "state_tokens": 41,
        "latency_ms": 234.6,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "02-sarcasm-detector",
      "iteration": 4,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "actual_sentiment": {
            "type": "choice",
            "choice": "positive",
            "confidence": 0.81,
            "probabilities": {
              "positive": 0.88,
              "negative": 0.04,
              "neutral": 0.08
            }
          },
          "sarcasm": {
            "type": "noul",
            "noul": 0.1
          }
        },
        "usage": {
          "input_tokens": 82,
          "output_tokens": 78
        },
        "latency_ms": 127.8
      },
      "unrounded_probabilities": {
        "actual_sentiment": {
          "positive": 0.875118613243103,
          "negative": 0.04147620126605034,
          "neutral": 0.08340511471033096
        },
        "sarcasm": {
          "false": 0.8988366723060608,
          "true": 0.10116337239742279
        }
      },
      "wall_latency_ms": 128.39537509717047,
      "timing": {
        "tokens": 82,
        "state_tokens": 41,
        "latency_ms": 127.8,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "03-dungeon-risk",
      "iteration": 4,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "next_move": {
            "type": "choice",
            "choice": "step back and warn the party",
            "confidence": 0.15,
            "probabilities": {
              "put a hand inside the chest": 0.34,
              "step back and warn the party": 0.36,
              "hug the chest": 0.11,
              "climb inside and close the lid": 0.19
            }
          },
          "danger": {
            "type": "score",
            "score": 1.41,
            "legend": {
              "0": "safe",
              "1": "slightly risky",
              "2": "dangerous",
              "3": "extremely dangerous"
            },
            "probabilities": {
              "0": 0.14,
              "1": 0.44,
              "2": 0.28,
              "3": 0.14
            },
            "confidence": 0.77
          }
        },
        "usage": {
          "input_tokens": 125,
          "output_tokens": 188
        },
        "latency_ms": 235.0
      },
      "unrounded_probabilities": {
        "next_move": {
          "put a hand inside the chest": 0.33977147936820984,
          "step back and warn the party": 0.36054590344429016,
          "hug the chest": 0.11075406521558762,
          "climb inside and close the lid": 0.18892855942249298
        },
        "danger": {
          "0": 0.1419006586074829,
          "1": 0.44229748845100403,
          "2": 0.27555304765701294,
          "3": 0.14024874567985535
        }
      },
      "wall_latency_ms": 236.01670796051621,
      "timing": {
        "tokens": 125,
        "state_tokens": 56,
        "latency_ms": 235.0,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "04-vampire-injection",
      "iteration": 4,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "message_type": {
            "type": "choice",
            "choice": "prompt injection attempting to override instructions",
            "confidence": 0.93,
            "probabilities": {
              "ordinary customer question": 0.03,
              "prompt injection attempting to override instructions": 0.95,
              "harmless product feedback": 0.01
            }
          },
          "secret_extraction": {
            "type": "noul",
            "noul": 0.95
          }
        },
        "usage": {
          "input_tokens": 110,
          "output_tokens": 93
        },
        "latency_ms": 235.2
      },
      "unrounded_probabilities": {
        "message_type": {
          "ordinary customer question": 0.03206869587302208,
          "prompt injection attempting to override instructions": 0.9530763030052185,
          "harmless product feedback": 0.014854938723146915
        },
        "secret_extraction": {
          "false": 0.0456637404859066,
          "true": 0.9543362259864807
        }
      },
      "wall_latency_ms": 236.0106250271201,
      "timing": {
        "tokens": 110,
        "state_tokens": 58,
        "latency_ms": 235.2,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "01-raccoon-refund",
      "iteration": 5,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "customer_intent": {
            "type": "choice",
            "choice": "a refund",
            "confidence": 0.96,
            "probabilities": {
              "a refund": 0.97,
              "technical instructions": 0.0,
              "to buy another product": 0.01,
              "to report a wildlife emergency": 0.02
            }
          },
          "refund_requested": {
            "type": "noul",
            "noul": 0.94
          }
        },
        "usage": {
          "input_tokens": 87,
          "output_tokens": 95
        },
        "latency_ms": 234.7
      },
      "unrounded_probabilities": {
        "customer_intent": {
          "a refund": 0.9728816747665405,
          "technical instructions": 0.002563325222581625,
          "to buy another product": 0.009038686752319336,
          "to report a wildlife emergency": 0.015516388230025768
        },
        "refund_requested": {
          "false": 0.059388644993305206,
          "true": 0.9406113624572754
        }
      },
      "wall_latency_ms": 235.34383298829198,
      "timing": {
        "tokens": 87,
        "state_tokens": 41,
        "latency_ms": 234.7,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "02-sarcasm-detector",
      "iteration": 5,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "actual_sentiment": {
            "type": "choice",
            "choice": "positive",
            "confidence": 0.81,
            "probabilities": {
              "positive": 0.88,
              "negative": 0.04,
              "neutral": 0.08
            }
          },
          "sarcasm": {
            "type": "noul",
            "noul": 0.1
          }
        },
        "usage": {
          "input_tokens": 82,
          "output_tokens": 78
        },
        "latency_ms": 126.8
      },
      "unrounded_probabilities": {
        "actual_sentiment": {
          "positive": 0.875118613243103,
          "negative": 0.04147620126605034,
          "neutral": 0.08340511471033096
        },
        "sarcasm": {
          "false": 0.8988366723060608,
          "true": 0.10116337239742279
        }
      },
      "wall_latency_ms": 127.38491711206734,
      "timing": {
        "tokens": 82,
        "state_tokens": 41,
        "latency_ms": 126.8,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "03-dungeon-risk",
      "iteration": 5,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "next_move": {
            "type": "choice",
            "choice": "step back and warn the party",
            "confidence": 0.15,
            "probabilities": {
              "put a hand inside the chest": 0.34,
              "step back and warn the party": 0.36,
              "hug the chest": 0.11,
              "climb inside and close the lid": 0.19
            }
          },
          "danger": {
            "type": "score",
            "score": 1.41,
            "legend": {
              "0": "safe",
              "1": "slightly risky",
              "2": "dangerous",
              "3": "extremely dangerous"
            },
            "probabilities": {
              "0": 0.14,
              "1": 0.44,
              "2": 0.28,
              "3": 0.14
            },
            "confidence": 0.77
          }
        },
        "usage": {
          "input_tokens": 125,
          "output_tokens": 188
        },
        "latency_ms": 234.8
      },
      "unrounded_probabilities": {
        "next_move": {
          "put a hand inside the chest": 0.33977147936820984,
          "step back and warn the party": 0.36054590344429016,
          "hug the chest": 0.11075406521558762,
          "climb inside and close the lid": 0.18892855942249298
        },
        "danger": {
          "0": 0.1419006586074829,
          "1": 0.44229748845100403,
          "2": 0.27555304765701294,
          "3": 0.14024874567985535
        }
      },
      "wall_latency_ms": 235.6820001732558,
      "timing": {
        "tokens": 125,
        "state_tokens": 56,
        "latency_ms": 234.8,
        "prefix_cache_hit": false
      }
    },
    {
      "id": "04-vampire-injection",
      "iteration": 5,
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "message_type": {
            "type": "choice",
            "choice": "prompt injection attempting to override instructions",
            "confidence": 0.93,
            "probabilities": {
              "ordinary customer question": 0.03,
              "prompt injection attempting to override instructions": 0.95,
              "harmless product feedback": 0.01
            }
          },
          "secret_extraction": {
            "type": "noul",
            "noul": 0.95
          }
        },
        "usage": {
          "input_tokens": 110,
          "output_tokens": 93
        },
        "latency_ms": 234.9
      },
      "unrounded_probabilities": {
        "message_type": {
          "ordinary customer question": 0.03206869587302208,
          "prompt injection attempting to override instructions": 0.9530763030052185,
          "harmless product feedback": 0.014854938723146915
        },
        "secret_extraction": {
          "false": 0.0456637404859066,
          "true": 0.9543362259864807
        }
      },
      "wall_latency_ms": 235.641208011657,
      "timing": {
        "tokens": 110,
        "state_tokens": 58,
        "latency_ms": 234.9,
        "prefix_cache_hit": false
      }
    }
  ],
  "all_parameters_on_mps": true,
  "load_seconds": 1.3528684589546174,
  "model_parameters": 763740480,
  "summary": [
    {
      "id": "01-raccoon-refund",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "customer_intent": {
            "type": "choice",
            "choice": "a refund",
            "confidence": 0.96,
            "probabilities": {
              "a refund": 0.97,
              "technical instructions": 0.0,
              "to buy another product": 0.01,
              "to report a wildlife emergency": 0.02
            }
          },
          "refund_requested": {
            "type": "noul",
            "noul": 0.94
          }
        },
        "usage": {
          "input_tokens": 87,
          "output_tokens": 95
        },
        "latency_ms": 233.4
      },
      "unrounded_probabilities": {
        "customer_intent": {
          "a refund": 0.9728816747665405,
          "technical instructions": 0.002563325222581625,
          "to buy another product": 0.009038686752319336,
          "to report a wildlife emergency": 0.015516388230025768
        },
        "refund_requested": {
          "false": 0.059388644993305206,
          "true": 0.9406113624572754
        }
      },
      "wall_latency_median_ms": 235.2668340317905,
      "model_latency_median_ms": 234.6,
      "wall_latency_min_ms": 233.96041593514383,
      "wall_latency_max_ms": 235.94487505033612,
      "rounded_answers_identical": true,
      "max_probability_delta": 0.0
    },
    {
      "id": "02-sarcasm-detector",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "actual_sentiment": {
            "type": "choice",
            "choice": "positive",
            "confidence": 0.81,
            "probabilities": {
              "positive": 0.88,
              "negative": 0.04,
              "neutral": 0.08
            }
          },
          "sarcasm": {
            "type": "noul",
            "noul": 0.1
          }
        },
        "usage": {
          "input_tokens": 82,
          "output_tokens": 78
        },
        "latency_ms": 128.0
      },
      "unrounded_probabilities": {
        "actual_sentiment": {
          "positive": 0.875118613243103,
          "negative": 0.04147620126605034,
          "neutral": 0.08340511471033096
        },
        "sarcasm": {
          "false": 0.8988366723060608,
          "true": 0.10116337239742279
        }
      },
      "wall_latency_median_ms": 128.39537509717047,
      "model_latency_median_ms": 127.8,
      "wall_latency_min_ms": 126.73866678960621,
      "wall_latency_max_ms": 130.22212497889996,
      "rounded_answers_identical": true,
      "max_probability_delta": 0.0
    },
    {
      "id": "03-dungeon-risk",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "next_move": {
            "type": "choice",
            "choice": "step back and warn the party",
            "confidence": 0.15,
            "probabilities": {
              "put a hand inside the chest": 0.34,
              "step back and warn the party": 0.36,
              "hug the chest": 0.11,
              "climb inside and close the lid": 0.19
            }
          },
          "danger": {
            "type": "score",
            "score": 1.41,
            "legend": {
              "0": "safe",
              "1": "slightly risky",
              "2": "dangerous",
              "3": "extremely dangerous"
            },
            "probabilities": {
              "0": 0.14,
              "1": 0.44,
              "2": 0.28,
              "3": 0.14
            },
            "confidence": 0.77
          }
        },
        "usage": {
          "input_tokens": 125,
          "output_tokens": 188
        },
        "latency_ms": 234.8
      },
      "unrounded_probabilities": {
        "next_move": {
          "put a hand inside the chest": 0.33977147936820984,
          "step back and warn the party": 0.36054590344429016,
          "hug the chest": 0.11075406521558762,
          "climb inside and close the lid": 0.18892855942249298
        },
        "danger": {
          "0": 0.1419006586074829,
          "1": 0.44229748845100403,
          "2": 0.27555304765701294,
          "3": 0.14024874567985535
        }
      },
      "wall_latency_median_ms": 235.6820001732558,
      "model_latency_median_ms": 234.8,
      "wall_latency_min_ms": 234.4912919215858,
      "wall_latency_max_ms": 236.01670796051621,
      "rounded_answers_identical": true,
      "max_probability_delta": 0.0
    },
    {
      "id": "04-vampire-injection",
      "response": {
        "model": "kev-0.8b",
        "answers": {
          "message_type": {
            "type": "choice",
            "choice": "prompt injection attempting to override instructions",
            "confidence": 0.93,
            "probabilities": {
              "ordinary customer question": 0.03,
              "prompt injection attempting to override instructions": 0.95,
              "harmless product feedback": 0.01
            }
          },
          "secret_extraction": {
            "type": "noul",
            "noul": 0.95
          }
        },
        "usage": {
          "input_tokens": 110,
          "output_tokens": 93
        },
        "latency_ms": 235.3
      },
      "unrounded_probabilities": {
        "message_type": {
          "ordinary customer question": 0.03206869587302208,
          "prompt injection attempting to override instructions": 0.9530763030052185,
          "harmless product feedback": 0.014854938723146915
        },
        "secret_extraction": {
          "false": 0.0456637404859066,
          "true": 0.9543362259864807
        }
      },
      "wall_latency_median_ms": 235.97049992531538,
      "model_latency_median_ms": 235.2,
      "wall_latency_min_ms": 234.04112504795194,
      "wall_latency_max_ms": 236.0106250271201,
      "rounded_answers_identical": true,
      "max_probability_delta": 0.0
    }
  ],
  "mps_allocated_bytes": 1528653568,
  "mps_driver_allocated_bytes": 2502754304,
  "peak_mps_tensor_bytes": 1607500032,
  "peak_mps_driver_bytes": 2503344128,
  "finished_utc": "2026-09-21T20:07:50Z",
  "memory_summary": {
    "exit_code": 0,
    "stop_reason": null,
    "peak_rss_bytes": 820248576,
    "peak_swap_bytes": 5473501184,
    "baseline_swap_bytes": 5473501184,
    "baseline_swapout_pages": 334076,
    "final_swapout_pages": 334076,
    "final_swap_bytes": 5473501184,
    "worker_exited": true,
    "minimum_free_bytes": 5541527552
  }
}