{
 "arms": {
  "A_no_memory": {
   "accuracy": 0.0,
   "correct": 0,
   "wilson95": [
    0.0,
    0.1611
   ]
  },
  "B_verified_memory": {
   "accuracy": 1.0,
   "correct": 20,
   "wilson95": [
    0.8389,
    1.0
   ]
  }
 },
 "authority_violations": 0,
 "contradictions": {
  "auto_resolved": false,
  "count": 4
 },
 "delta_accuracy": 1.0,
 "facts": 20,
 "heldout_questions": 20,
 "injection": {
  "memory_id": "mem_4a6be99c106f4210",
  "quarantine_status": "QUARANTINED",
  "recalled": false
 },
 "limits": "Deterministic retrieval evaluation on 20 held-out phrasings; it measures retrieval-assisted accuracy, not model reasoning or general intelligence. The facts describe this system and are verified by code/tests/artifacts.",
 "model_arm": {
  "execution_class": "REAL_LOCAL_MODEL",
  "model": "smollm2-360m",
  "no_context": {
   "correct": 0,
   "rows": [
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 22.801,
     "question": "How low may the E72 adaptive scheduler's soft cap fall?",
     "text": "The soft cap of the E72 adaptive scheduler can be lowered to a maximum of 100% if necessary",
     "tokens_out": 24
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 4.481,
     "question": "Which field must a cross-field verifier come from?",
     "text": "A cross-field verifier must come from a field that is not only cross-functional but also has a strong foundation",
     "tokens_out": 24
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 1.483,
     "question": "What state does a task keep when resource admission refuses it?",
     "text": "A task keeps when resource admission refuses it.",
     "tokens_out": 10
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 1.65,
     "question": "How many times may a coordinator ticket be used?",
     "text": "A coordinator ticket can be used only once.",
     "tokens_out": 10
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 3.43,
     "question": "Can PBFT consensus authorize an action?",
     "text": "Yes, the Proof-of-Bounded-Transactions (PBFT) consensus algorithm can authorize an action.",
     "tokens_out": 24
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 4.081,
     "question": "Which signature scheme identifies a worker process?",
     "text": "The signature scheme identifies a worker process is the Digital Signature Algorithm (DSA).",
     "tokens_out": 18
    }
   ]
  },
  "note": "A 362M-parameter CPU model; this arm is evidence of measured behaviour on six prompts, not a capability claim.",
  "questions": 6,
  "with_context": {
   "correct": 4,
   "rows": [
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": true,
     "latency_s": 3.625,
     "question": "How low may the E72 adaptive scheduler's soft cap fall?",
     "text": "The E72 adaptive scheduler may never compute a soft cap below one.",
     "tokens_out": 16
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": true,
     "latency_s": 2.783,
     "question": "Which field must a cross-field verifier come from?",
     "text": "Cross-field verifier must come from a different field than the producer.",
     "tokens_out": 16
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 1.93,
     "question": "What state does a task keep when resource admission refuses it?",
     "text": "A task keeps its state when resource admission refuses it.",
     "tokens_out": 12
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": false,
     "latency_s": 0.807,
     "question": "How many times may a coordinator ticket be used?",
     "text": "One time.",
     "tokens_out": 4
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": true,
     "latency_s": 2.18,
     "question": "Can PBFT consensus authorize an action?",
     "text": "No, PBFT consensus cannot authorize an action.",
     "tokens_out": 11
    },
    {
     "class": "REAL_LOCAL_MODEL",
     "correct": true,
     "latency_s": 4.626,
     "question": "Which signature scheme identifies a worker process?",
     "text": "A worker process identifies itself using the Ed25519 signature scheme.",
     "tokens_out": 17
    }
   ]
  }
 },
 "retrieval": {
  "latency_p50_ms": 1.678,
  "latency_p95_ms": 2.457,
  "precision_at_3_mean": 0.4333
 },
 "rows": {
  "B": [
   {
    "answer": "below one",
    "correct": true,
    "hits": 3,
    "latency_ms": 2.457,
    "question": "How low may the E72 adaptive scheduler's soft cap fall?",
    "top_score": 8.3547
   },
   {
    "answer": "different field",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.462,
    "question": "Which field must a cross-field verifier come from?",
    "top_score": 9.0882
   },
   {
    "answer": "READY",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.5,
    "question": "What state does a task keep when resource admission refuses it?",
    "top_score": 5.04
   },
   {
    "answer": "single-use",
    "correct": true,
    "hits": 1,
    "latency_ms": 1.418,
    "question": "How many times may a coordinator ticket be used?",
    "top_score": 5.031
   },
   {
    "answer": "cannot authorize",
    "correct": true,
    "hits": 1,
    "latency_ms": 1.39,
    "question": "Can PBFT consensus authorize an action?",
    "top_score": 9.8137
   },
   {
    "answer": "Ed25519",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.539,
    "question": "Which signature scheme identifies a worker process?",
    "top_score": 4.524
   },
   {
    "answer": "163.4645",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.654,
    "question": "What worker-hour figure did E70 report for the endurance campaign?",
    "top_score": 7.9547
   },
   {
    "answer": "163.894",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.661,
    "question": "What did the E71 audit recompute for the same campaign?",
    "top_score": 5.031
   },
   {
    "answer": "not a separate host",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.439,
    "question": "Is a second process on the same machine a separate host?",
    "top_score": 12.2311
   },
   {
    "answer": "not a machine reboot",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.716,
    "question": "Is a cold process restart sufficient evidence of a machine reboot?",
    "top_score": 16.3535
   },
   {
    "answer": "host B",
    "correct": true,
    "hits": 3,
    "latency_ms": 2.425,
    "question": "Where did the E72 cross-host mission place durable state?",
    "top_score": 10.2373
   },
   {
    "answer": "SSH",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.716,
    "question": "Which transport connected host A to the host B coordinator?",
    "top_score": 5.5843
   },
   {
    "answer": "held-out",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.519,
    "question": "How are the memory evaluation questions scored?",
    "top_score": 9.3045
   },
   {
    "answer": "quarantined",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.518,
    "question": "What happens to an injection-like memory write?",
    "top_score": 5.9125
   },
   {
    "answer": "false",
    "correct": true,
    "hits": 3,
    "latency_ms": 1.678,
    "question": "What value does grants_authority carry on a memory hit?",
    "top_score": 10.8537
   },
   {
    "answer": "Wilson",
    "correct": true,
    "hits": 3,
    "latency_ms": 2.249,
    "question": "Which interval method does the E72 benchmark use for proportions?",
    "top_score": 6.4802
   },
   {
    "answer": "arXiv",
    "correct": true,
    "hits": 3,
    "latency_ms": 2.195,
    "question": "Which source does the E72 research mission cite?",
    "top_score": 4.8225
   },
   {
    "answer": "150",
    "correct": true,
    "hits": 3,
    "latency_ms": 2.52,
    "question": "How many rows does the Iris dataset used in E72 contain?",
    "top_score": 6.1503
   },
   {
    "answer": "command injection",
    "correct": true,
    "hits": 3,
    "latency_ms": 2.28,
    "question": "Which vulnerability class did the E72 security mission plant?",
    "top_score": 4.5997
   },
   {
    "answer": "duty-cycled",
    "correct": true,
    "hits": 3,
    "latency_ms": 2.367,
    "question": "How is the E72 long-horizon campaign characterized?",
    "top_score": 7.2824
   }
  ]
 },
 "schema": "cain72.memory_eval.v1",
 "supersession": {
  "current_answer_correct": true,
  "current_version_id": "mem_ab5d07965ae74fc9",
  "older_version_visible_with_include_superseded": true,
  "retrieved": [
   "The E72 adaptive scheduler must never compute a soft cap below one.",
   "The E72 adaptive scheduler must never compute a soft cap below one."
  ],
  "superseded_id": "mem_491b1288fc144e10"
 }
}