# Study 73: Human Conversation Corpora & Deterministic Piecewise-Affine (PWA) Benchmark Court ($\mathcal{M}_8 = S^4 \times C^4$)

**Executable Court:** [`reproduce/study-73-human-conversation-corpora-court.swift`](reproduce/study-73-human-conversation-corpora-court.swift)  
**Live Feed:** [`/feeds/human-conversation-benchmarks.json`](cells/xcode/Resources/language-game/feeds/human-conversation-benchmarks.json)  
**Verified Seal:** `FNV-1a64: c37e45fba5539ebf` · `no_float: 1`

---

## 1. Six Gold-Standard Human Conversation Corpora on $\mathcal{M}_8 = S^4 \times C^4$

Calibrating a deterministic Piecewise-Affine (PWA) conversational engine—and hardening our detector against rogue autonomous chatbots—requires annotated human pragmatics corpora and state-transition dialogue benchmarks that measure microsecond turn-taking latency ($\Delta \tau$), speech-act state transitions (DAMSL / ISO 24617-2), conversational repair loops, and emotional arousal damping ($-\Delta \alpha$).

| Corpus / Dataset | Scale & Modality | Native Structural Annotations | Direct Mapping to $\mathcal{M}_8 = S^4 \times C^4$ Lattice |
| --- | --- | --- | --- |
| **Switchboard Dialog Act Corpus (SwDA) + NXT Timings** | 1,155 human telephone conversations (~205,000 utterances) | Tagged with 42 DAMSL Speech-Act primitives (`Statement`, `Yes-No-Question`, `Backchannel`, `Hedge`, `Apology`, `Collaborative Completion`) + word-level millisecond timestamps. | **Baseline Human Cadence ($\tau \in S^4$):** Provides the exact distribution of human turn-taking gaps ($150\text{–}350\text{ ms}$) and backchannel frequency to separate real humans from GPU batch-scheduled bots. |
| **DailyDialog** | 13,118 multi-turn everyday human dialogues | Every turn is dual-labeled for **Communicative Intent** (`Inform`, `Question`, `Directive`, `Commissive`) and **Emotion State** (Ekman 6 + Neutral). | **Everyday State-Transition Matrix ($\mathbf{M}_{\text{chat}}$):** Maps how ordinary human small talk, scheduling, and polite disagreement transition between $(\alpha_k, \rho_k, \iota_k)$ states. |
| **ESConv (Emotional Support Conversation) & EmpatheticDialogues** | 1,300+ long-horizon crisis/support chats + 25,000 empathetic turns | Annotated with 8 Helping/De-escalation Strategies (`Question`, `Restatement/Paraphrasing`, `Reflection of Feelings`, `Self-Disclosure`, `Affirmation`, `Providing Suggestions`) and User Distress Score ($\alpha_k$: 1 to 5) before and after. | **Direct LEAPS / Stairway Calibration:** Measures the exact integer drop in arousal ($-\Delta \alpha$) produced by each de-escalation language game. |
| **PersuasionForGood & CraigslistBargains** | 1,000+ asymmetric negotiation & persuasion dialogues | Turn-by-turn annotation of 54 persuasion/manipulation tactics (urgency framing, social proof, emotional appeal, foot-in-the-door) vs. buyer resistance. | **Adversarial Scammer / Bot Benchmark:** Calibrates the guard hyperplane $H_{\text{scam}} = \alpha_k - \iota_k - 200 \ge 0$ to detect predatory extraction games (pig-butchering / phishing). |
| **MultiWOZ 2.4 & SGD (Schema-Guided Dialogue)** | 30,000+ task-oriented human dialogues across 20+ domains | Exact Slot-Value State Tracking (belief states, user goals, system actions, API call triggers) at every turn. | **Zero-Overkill Task Completion ($\pmod x$):** Tests whether the chat reaches exact slot-lock ($\iota_k \to 1000$) in minimum turns without redundant questions. |
| **Conversation Gone Awry (ConvoKit — Cornell)** | Thousands of Wikipedia talk-page & Reddit threads paired with clean controls | Identifies threads that start civil ($\alpha_0 < 200$) and cross the Toxicity / Personal Attack Cliff ($H_{\text{kinetic}} \ge 0$) within 3–6 turns. | **Early Hyperplane Breach Benchmark:** Verifies that the Fiedler/Affine guard detects an impending conversational derailment $\ge 3$ turns before the toxicity cliff. |

---

## 2. The Five Mathematical Invariants of Natural Human Conversation

1. **Stivers' Universal Turn-Taking Invariant ($\Delta \tau \approx 200\text{ ms}$):**
   - Across all human languages (Stivers et al., *PNAS*), the median gap between Turn $A$ ending and Turn $B$ starting is $100\text{ to }300\text{ ms}$ (`Affine median = 205 ms, jitter = 69 ms`).
   - Cloud LLM pipelines exhibit a bimodal signature: a dead $800\text{–}2500\text{ ms}$ Time-To-First-Token (TTFT) followed by rigid $20\text{ ms}$ GPU token ticks (`Cloud LLM median = 20 ms -> BOT_FLAGGED`).
2. **Gricean Relevance & Slot-Packing Proportionality (`proportionalityScore = 1000/1000`):**
   - When an interlocutor sends a casual greeting or short check-in (`DialogueActPrimitive.greetingOrPhatic`, e.g., `"Hey, you around?"` or `"are you there"`), a natural human replies in $\le 25$ words (`"Yeah, right here—what's on your mind?"` or `"Yes, I'm right here with you! How is your day going, and what's on your mind?"` = 16 words).
   - Standard RLHF chatbots dump a 120-word bulleted wall of text (`Cloud LLM = 600/1000 -> FAIL_VERBOSE`).
3. **Other-Initiated Conversational Repair Elasticity (`repairElasticityScore = 1000/1000`):**
   - Humans clarify ambiguity in a compact 3-turn micro-loop (`"Wait, the Tuesday one?"` $\to$ `"Yeah, the 9 AM."` $\to$ `"Got it."`) that resolves $\Delta \iota > 0$ within $\le 2$ turns.
4. **Affective Damping (Asymmetric $\mathbf{M}_{\text{empathy}}$, `affectDampingScore = 1000/1000`):**
   - When Speaker $A$ enters at high arousal ($\alpha_A > 600$), Speaker $B$ executes an `affectValidation` or `boundedProposal` act at low arousal ($\alpha_B < 250$), pulling $\alpha_A$ down on the next turn ($\alpha_A: 720 \to 410 \to 220$).
5. **Bounded Persona Invariance (Zero Sycophancy Drift & Early $H_{\text{scam}}$ Detection):**
   - When a scammer injects artificial urgency ($\Delta \alpha \gg 0$) while withholding verifiable grounding ($\Delta \iota \le 0$), the guard hyperplane $H_{\text{scam}} = \alpha_k - \iota_k - 200 \ge 0$ trips at Turn 2—**3 full turns** before the conversation crosses the kinetic cliff at Turn 5.