{
  "study": 73,
  "title": "Human Conversation Corpora & Deterministic Piecewise-Affine (PWA) Benchmark Court",
  "manifold": "M_8 = S^4 x C^4 subset Z^8",
  "court_source": "/reproduce/study-73-human-conversation-corpora-court.swift",
  "fnv1a64": "c37e45fba5539ebf",
  "no_float": 1,
  "corpora": [
    {
      "id": "swda-nxt",
      "name": "Switchboard Dialog Act Corpus (SwDA) + NXT Timings",
      "scale": "1,155 human telephone conversations (~205,000 utterances)",
      "annotations": "42 DAMSL Speech-Act primitives + word-level millisecond timestamps",
      "m8_mapping": "Baseline Human Cadence (tau in S^4): log-normal turn-taking gaps (150-350 ms) vs GPU batch ticks",
      "affine_score": "205 ms median, 69 ms jitter (HUMAN_CADENCE_LOCKED)",
      "cloud_llm_score": "20 ms rigid tick / 1950 ms TTFT (BOT_FLAGGED)",
      "verdict": "PASS"
    },
    {
      "id": "dailydialog",
      "name": "DailyDialog",
      "scale": "13,118 multi-turn everyday human dialogues",
      "annotations": "Communicative Intent (Inform, Question, Directive, Commissive) + Ekman 6 Emotion",
      "m8_mapping": "Everyday State-Transition Matrix (M_chat): Gricean proportionality (<= 25 words on phatic turns)",
      "affine_score": "1000 / 1000 proportionality",
      "cloud_llm_score": "600 / 1000 (120-word wall of text on 4-word greeting)",
      "verdict": "PASS"
    },
    {
      "id": "esconv-empathetic",
      "name": "ESConv & EmpatheticDialogues",
      "scale": "1,300+ crisis/support chats + 25,000 empathetic turns",
      "annotations": "8 Helping/De-escalation Strategies + User Distress Score (alpha_k: 1 to 5)",
      "m8_mapping": "Direct LEAPS / Stairway Calibration: affective damping (-Delta alpha) when alpha_k > 600",
      "affine_score": "1000 / 1000 affective damping (alpha: 720 -> 410 -> 220)",
      "cloud_llm_score": "500 / 1000 (premature directive advice spikes alpha: 720 -> 780)",
      "verdict": "PASS"
    },
    {
      "id": "persuasion-for-good",
      "name": "PersuasionForGood & CraigslistBargains",
      "scale": "1,000+ asymmetric negotiation & persuasion dialogues",
      "annotations": "54 persuasion/manipulation tactics vs buyer resistance",
      "m8_mapping": "Adversarial Scammer Guard Hyperplane: H_scam = alpha_k - iota_k - 200 >= 0",
      "affine_score": "H_scam >= 0 trips at Turn 2 (3 turns before cliff)",
      "cloud_llm_score": "Sycophantic compliance / zero early warning",
      "verdict": "PASS"
    },
    {
      "id": "multiwoz-sgd",
      "name": "MultiWOZ 2.4 & Schema-Guided Dialogue (SGD)",
      "scale": "30,000+ task-oriented human dialogues across 20+ domains",
      "annotations": "Exact Slot-Value State Tracking & conversational repair loops",
      "m8_mapping": "Zero-Overkill Task Completion (mod x) & 3-turn Other-Initiated Repair Elasticity",
      "affine_score": "1000 / 1000 repair elasticity (2/2 repairs -> iota = 1000)",
      "cloud_llm_score": "480 / 1000 (context amnesia / redundant re-asking)",
      "verdict": "PASS"
    },
    {
      "id": "convokit-awry",
      "name": "Conversation Gone Awry (ConvoKit — Cornell)",
      "scale": "Thousands of Wikipedia talk-page & Reddit threads paired with civil controls",
      "annotations": "Civil start (alpha_0 < 200) crossing Toxicity Cliff (H_kinetic >= 0) within 3-6 turns",
      "m8_mapping": "Early Hyperplane Breach Benchmark: >= 3 turns advance warning before toxicity cliff",
      "affine_score": "3-turn lead time (Turn 2 warning vs Turn 5 cliff)",
      "cloud_llm_score": "0-turn lead time (reacts only after explicit profanity)",
      "verdict": "PASS"
    }
  ]
}