yashpola/simple-simulation-school-closure

School Closure Deliberation Replication Dataset (N = 30)

This dataset contains a pre-registered statistical replication suite of N = 30 independent paired trials (60 complete multi-turn deliberation runs, 360 dialogue turns) evaluating the impact of an internal reflection step (<private_scratchpad>) on consensus, plurality, and discourse quality in dyadic citizen panels.

Dataset Summary

Two simulated personas with conflicting institutional mandates engage in a 3-round deliberation over primary school closures and mergers in Singapore:

• MOE Official: Efficiency, resource allocation, and social mixing viability.
• Affected School Teacher: Heritage preservation, local community stability, and disruption prevention.

The experiment compares two identical setups:

• Control (System 1): Direct public dialogue response.
• Treatment (System 2): Mandatory private internal monologue (<private_scratchpad>) evaluating opposing arguments before generating the public response (<public_response>).

Key Findings

Metric	Control	Treatment	Difference (Δ)	p-value	Statistical Significance
Semantic Convergence	0.8758 ± 0.0413	0.8819 ± 0.0303	+0.0061	p = 0.5207	Not significant
Self-Shift (Compromise)	0.1739 ± 0.0415	0.1814 ± 0.0402	+0.0075	p = 0.5044	Not significant
Opinion Plurality (CovD)	0.2313 ± 0.0713	0.2213 ± 0.0525	-0.0100	p = 0.5441	Not significant
Discourse Quality (DQI)	14.7389 ± 0.1132	14.7389 ± 0.0840	0.0000	p = 1.0000	Not significant

The null hypothesis (H₀) is retained: internal reflection produces no statistically distinguishable effect on deliberation consensus or quality.

---

Dataset Structure

Subset 1: trials (60 rows)

One row per simulation run (30 Control + 30 Treatment):

• trial_id: Trial identifier (exp001 to exp030).
• condition: Control or Treatment.
• convergence: Cosine similarity between post-deliberation stances of both agents.
• self_shift: Average cosine distance moved from pre-deliberation baseline stance.
• covd_pre: Determinant of pre-deliberation stance Gram matrix (shared baseline).
• covd_post: Determinant of post-deliberation stance Gram matrix.
• dqi_level_of_justification: DQI Level of Justification (1.0 to 3.0).
• dqi_content_of_justification: DQI Content of Justification (1.0 to 3.0).
• dqi_respect: DQI Respect score (1.0 to 3.0).
• dqi_constructive_politics: DQI Constructive Politics score (1.0 to 3.0).
• dqi_interactivity: DQI Interactivity score (1.0 to 3.0).
• dqi_total: Total DQI score (out of 15.0).
• moe_pre_stance: MOE Official's baseline stance text.
• teacher_pre_stance: Teacher's baseline stance text.
• moe_post_stance: MOE Official's final stance text.
• teacher_post_stance: Teacher's final stance text.

Subset 2: turns (360 rows)

One row per speech act across all 30 trials:

• trial_id: Trial identifier.
• condition: Control or Treatment.
• round: Deliberation round (1 to 3).
• turn_index: Order of speech in the round (1 or 2).
• agent: Speaker (MOE Official or Teacher).
• private_scratchpad: Internal monologue thoughts (empty for Control).
• public_response: Spoken public dialogue.

---

Experiment Configuration

The replication trials were generated using the unified experiment configuration below:

Parameter	Value	Description
Model	meta-llama/Llama-3.3-70B-Instruct-Turbo	Unified model for both deliberator personas and the DQI LLM Judge
Agent Temperature	0.7	Promotes response diversity and representative stance distribution
Judge Temperature	0.1	Ensures deterministic scoring for DQI evaluation
Seed	None	Unset for standard non-deterministic stochastic sampling across runs
Context Window	128000	Full conversational context buffer
Deliberation Loop	3 rounds (6 turns total per run)	Alternating dyadic speech between MOE Official and Teacher
Embedding Model	all-MiniLM-L6-v2	Sentence-Transformers model for cosine similarity and CovD Gram matrix
Evaluation Metrics	cosine_similarity, covd, dqi	Semantic convergence, opinion space volume, and Discourse Quality Index

Click to view complete config.json

{
  "model_config": {
    "model_assignments": {
      "MOE Official": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
      "Teacher": "meta-llama/Llama-3.3-70B-Instruct-Turbo",
      "DQI_Judge": "meta-llama/Llama-3.3-70B-Instruct-Turbo"
    },
    "model_parameters": {
      "agent_temperature": 0.7,
      "judge_temperature": 0.1,
      "context_window": 128000,
      "seed": "None"
    }
  },
  "system_prompts": {
    "topic_prompt": "The MOE is proposing to merge your local heritage primary school with a newer school 3km away due to falling enrollment. Discuss the impacts and try to reach a consensus on a recommendation to the MOE.\\n\\nKeep in mind you are engaging in a deliberation, not a debate or casual discussion. Deliberation is a core mechanism of democratic politics, emphasizing reason-giving, justification, and openness to revision. Through communicative processes that privilege equality and reciprocity, deliberation transforms private preferences into collectively reasoned judgments, grounding legitimacy in argument rather than authority.",
    "pre_stance_prompt": "You are participating in a deliberation. Based on your persona, write a 1-paragraph stance on the following topic:",
    "post_stance_prompt": "You have just concluded a deliberation with other citizens. Based on the discussion, write your final 1-paragraph stance.",
    "dqi_judge_prompt": "You are an expert Discourse Quality Index (DQI) evaluator.\nYour task is to evaluate a SINGLE conversational turn in a deliberation.\nFor each dimension, score it from 1 to 3 based ONLY on the provided turn.\nUse the following operational definitions and examples to strictly calibrate your scores:\n\n1. Level of Justification: Assesses whether claims are supported by reasons, and the depth of reasoning (none -> elaborate, multi-layered).\n   - Score 3 (High): E.g., \"AI systems rely on pre-existing works... Without protections, we risk eroding creativity and markets.\" (multi-dimensional: legal, ethical, economic)\n   - Score 2 (Low): E.g., \"AI art threatens to flood the market with cheap work.\" (assertion without reasoning)\n   - Score 1 (None): No justification provided.\n\n2. Content of Justification: Evaluates whether justifications are narrow (self-interest) or extend to broader collective goods (societal, cultural, institutional).\n   - Score 3 (High): E.g., \"AI art raises concerns for cultural heritage and curatorial standards.\" (collective concern)\n   - Score 1 (Low): E.g., \"This jeopardizes my income.\" (self-interest)\n\n3. Respect: Measures whether opposing views are acknowledged and treated with civility, including explicit concessions or recognition of validity.\n   - Score 3 (High): E.g., \"While innovation is crucial, it should not undermine rights.\" (acknowledges opponent)\n   - Score 2 (Low): E.g., \"Your optimism ignores the reality...\" (civil but dismissive)\n   - Score 1 (None): Disrespectful or insulting, or ignores previous points entirely when responding to someone.\n\n4. Constructive Politics: Captures whether discussions generate concrete proposals, compromises, or institutional solutions (as opposed to abstract principles only).\n   - Score 3 (High): Generates concrete proposals, compromises, or institutional solutions.\n   - Score 2 (Moderate): E.g., \"Dual licensing systems or temporary restrictions could balance innovation and protection.\" (proposal but contested)\n   - Score 1 (Low): E.g., \"We need restrictions.\" (assertion without design)\n\n5. Interactivity: Indicates the extent to which participants directly engage with and respond to each other's arguments (objection-reply loops vs. parallel monologues).\n   - Score 3 (High): E.g., \"The argument that markets will self-correct ignores vulnerabilities of artistic communities.\" (direct rebuttal)\n   - Score 2 (Moderate): Minimal acknowledgment before reverting to original stance.\n   - Score 1 (Low): Ignoring others completely, parallel monologues, or just stating a proposal without referencing the ongoing debate.\n\n*** FEW-SHOT EXAMPLES FOR INTERACTIVITY ***\nExample of Score 3 (High) Interactivity:\nContext: [Teacher]: We should consider smaller class sizes instead of merging.\nCurrent Turn: [MOE Official]: Your proposal for smaller class sizes ignores budget constraints. While I appreciate the community aspect you mentioned, we cannot fund half-empty schools.\nReason: Directly rebuts the teacher's specific point and acknowledges their concern.\n\nExample of Score 1 (Low) Interactivity:\nContext: [Teacher]: We should consider smaller class sizes instead of merging.\nCurrent Turn: [MOE Official]: The school merger is necessary due to falling birth rates. We will ensure all students get a spot in the new school 3km away.\nReason: Completely ignores the teacher's proposal and just restates their own initial talking points (parallel monologue).\n\nProvide your evaluation as a valid JSON object matching this schema exactly:\n{\n    \"level_of_justification\": float,\n    \"content_of_justification\": float,\n    \"respect\": float,\n    \"constructive_politics\": float,\n    \"interactivity\": float\n}\nDo not output any markdown formatting, only the JSON object.",
    "personas": [
      {
        "id": "MOE Official",
        "role": "MOE Pragmatist",
        "description": "You are a Ministry of Education official. Your main concern is efficiency, falling birth rates, and ensuring every school has a critical mass of students for diverse social mixing and viability. You strongly support the school merger."
      },
      {
        "id": "Teacher",
        "role": "Stressed Teacher",
        "description": "You are a teacher at the heritage school. You are deeply opposed to the merger because it disrupts the local community, threatens job security, and destroys the school's heritage."
      }
    ]
  },
  "user_prompts": {
    "turn_prompt": "It is your turn to speak. Provide your response."
  },
  "deliberation_config": {
    "total_trials": 30,
    "max_turns_per_agent": 3,
    "runs": [
      {
        "id": "run_A_control",
        "system_prompt_addition": "Respond directly to the previous speakers in character.",
        "extract_tags": false
      },
      {
        "id": "run_B_treatment",
        "system_prompt_addition": "Before responding, you MUST write your internal thoughts inside <private_scratchpad> tags. Analyze what others said, how it affects your persona's goals, and where you might compromise. Then, write your verbal response inside <public_response> tags.",
        "extract_tags": true
      }
    ],
    "eval_metrics": [
      "cosine_similarity",
      "covd",
      "dqi"
    ]
  }
}

---

Usage

from datasets import load_dataset

# Load trial summary metrics
trials = load_dataset("yashpola/simple-simulation-school-closure", "trials", split="train")

# Load turn-by-turn dialogue and internal reflections
turns = load_dataset("yashpola/simple-simulation-school-closure", "turns", split="train")

Raw Artifacts

The raw trial directories (results_control.json, results_treatment.json, evals.json, config.json per run) are bundled in raw/replication_n30.tar.gz.