Dialog Synthesis

┌─────────────────┐
│ steering_node   │  Selects next speaker, mood shift, continuation
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│ character_node  │  Generates dialog turn with persona-derived params
└────────┬────────┘
         │
         ▼
┌─────────────────┐
│quality_gate_node│  Three-level evaluation + naturalness scoring
└─────────────────┘

# From entity tensor state → LLM API params
temperature = base_temp * (1 + arousal * 0.3)
top_p = base_top_p * (1 - confidence * 0.1)
max_tokens = int(base_tokens * energy_factor)
frequency_penalty = 0.3 * patience

"I understand your concern, and I want to validate your feelings about this."

"The data doesn't support that timeline. We need three more weeks."

"Let's circle back on this offline and touch base next week to move forward."

"I'll send the revised figures Thursday. We can decide then."

"This is an important conversation to have, and I think we're making progress."

"The budget's tight, but we can make it work if engineering cuts two features."

"We should consider various approaches and evaluate our options going forward."

"Deploy to 10% of users Friday. If crash rate stays under 0.1%, full rollout Monday."

"I do not think we are ready. We will need more time."

"We're not ready. Need another week, minimum."

"It's worth noting that there are several factors we need to take into account."

"Revenue's down 12%. We cut marketing or we miss payroll in October."

"Look, I get what you're saying, but—"
"The timeline's unrealistic. You know that."
"Three weeks. That's it. Take it or leave it."

{
  "argument_style": "data-first; cites specific measurements; uses conditional logic",
  "disagreement_pattern": "asks for the source; names the exact number they dispute",
  "deflection_style": "redirects to technical subproblem",
  "sentence_style": "short declarative sentences; technical vocabulary",
  "never_does": ["make emotional appeals", "appeal to authority without data"],
  "signature_moves": ["qualifies estimates with error margins"]
}

"The O2 scrubber reading was 847 ppm at 14:23. Threshold is 800. 
We're 6% over spec."

{
  "argument_style": "schedule and budget framing; translates to downstream impact",
  "disagreement_pattern": "reframes as resource problem; offers to table for future",
  "deflection_style": "elevates ('let's not get into the weeds')",
  "sentence_style": "longer compound sentences; management vocabulary",
  "never_does": ["admit uncertainty in front of subordinates"],
  "signature_moves": ["ends turns with action items or deadlines"]
}

"We'll schedule a deep dive for Q2, but right now we need to focus on 
the deliverables we committed to the board. Sarah, can you have the 
revised timeline to me by Thursday?"

{
  "argument_style": "chain of command framing; mission risk vs crew safety",
  "disagreement_pattern": "asks for options not problems",
  "deflection_style": "defers to protocol or asks for formal assessment",
  "sentence_style": "crisp, clipped sentences; active voice; minimal hedging",
  "never_does": ["show fear in front of crew"],
  "signature_moves": ["asks 'what are our options'"]
}

"Status report. Now. What are our options?"

{
  "argument_style": "hypothesis-driven; cites studies and precedent",
  "disagreement_pattern": "questions methodology; asks about sample size",
  "deflection_style": "requests more data before committing",
  "sentence_style": "precise language; hedged claims; avoids absolutes",
  "never_does": ["claim certainty without evidence"],
  "signature_moves": ["prefaces claims with confidence level"]
}

"The data suggests—with about 75% confidence—that the anomaly is 
statistical noise, not a systemic failure. We'd need three more samples 
to rule out contamination."

{
  "argument_style": "constituency framing; appeals to shared values",
  "disagreement_pattern": "pivots to adjacent issue; acknowledges concern",
  "deflection_style": "broadens scope ('the real question is...')",
  "sentence_style": "rhythmic phrasing; inclusive language",
  "never_does": ["give a simple yes or no"],
  "signature_moves": ["triangulates between factions"]
}

"Look, we all want the same thing here—a system that works for everyone. 
The question isn't whether we act, it's how we balance competing priorities 
while keeping faith with the people who sent us here."

"The data serves as a testament to the transformative potential of our 
monitoring systems, showcasing the intricate interplay between sensor 
readings and operational outcomes."

# Arousal affects temperature (more aroused = more varied)
temperature = base_temperature * (1 + emotional_arousal * 0.3)

# Fatigue affects output length
max_tokens = int(base_max_tokens * (energy_budget / 100))

# Confidence affects top_p sampling
top_p = base_top_p * (1 - decision_confidence * 0.1)

# Patience affects frequency penalty
frequency_penalty = 0.3 * (patience_threshold / 100)

# ADPRS phi scales all parameters
all_params *= adprs_phi_current

back_layer = {
    "true_emotional_state": "Frustrated by Webb's dismissal",
    "withheld_knowledge": ["O2 reading at 847 ppm", "Sensor calibration date overdue"],
    "suppressed_impulses": ["Want to escalate to mission control"],
    "relationship_tension": {"Webb": -0.3}
}

front_layer = {
    "known_facts": ["Mission timeline", "Crew roles", "Current status"],
    "relationships": ["Webb: commander (strained)", "Chen: colleague (trusted)"],
    "recent_events": ["Sensor alert 2 hours ago", "Disagreement with Webb"]
}

# At timepoint T3, character doesn't know about events at T5
if timepoint_id < knowledge_first_exposed_at:
    filtered_knowledge.remove(knowledge_item)

# Naive capitalization-based extraction
knowledge = [word for word in content.split() if word[0].isupper()]
# Result: ["We'll", "Thanks", "What"] ❌

from workflows.knowledge_extraction import extract_knowledge_from_dialog

knowledge_items = extract_knowledge_from_dialog(
    dialog_turns=turns,
    context={"entities": entities, "timepoint": timepoint}
)
# Result: [
#   {"content": "O2 scrubber threshold is 800 ppm", 
#    "category": "fact", "confidence": 0.9},
#   {"content": "Sensor calibration overdue", 
#    "category": "observation", "confidence": 0.7}
# ]

{
  "dialog_attempts": [
    {
      "timepoint_id": "T2",
      "tactic_used": "data_argument",
      "target_entity": "Webb",
      "outcome": "dismissed",
      "argument_summary": "Presented O2 readings showing..."
    }
  ],
  "trust_ledger": {
    "Webb": -0.15,  # Trust decreased after dismissal
    "Chen": 0.03    # Trust increased after support
  },
  "unspoken_accumulation": [
    {
      "content": "O2 fault reading",
      "urgency": 0.8,  # Grows each time it's not expressed
      "first_formed": "T1"
    }
  ]
}

# Compute dialog impact
dialog_impact = analyze_dialog_emotional_impact(turns, speaker_id)

# Apply arousal decay (prevents saturation)
arousal_baseline = 0.3
arousal_decay_rate = 0.15
decayed_arousal = baseline + (current - baseline) * (1 - decay_rate)

# Update entity
new_valence = current_valence + dialog_impact["valence_delta"]
new_arousal = decayed_arousal + dialog_impact["arousal_delta"]

# Clamp to valid ranges
new_valence = max(-1.0, min(1.0, new_valence))
new_arousal = max(0.0, min(1.0, new_arousal))

"entity_metadata": {
  "archetype_id": "engineer",
  "role": "Flight Engineer"
}

max_turns = 10  # Prevent endless conversations
energy_drain_per_turn = 0.02  # Adjust based on scenario intensity

quality_config = {
    "enable_semantic_evaluation": True,
    "naturalness_threshold": 7.0,
    "max_retries": 2
}

# Before dialog: TTM → Cognitive
_sync_ttm_to_cognitive(entity)

# After dialog: Cognitive → TTM (learning)
_sync_cognitive_to_ttm(entity, updated_cognitive, store)