⚠️ Status: Archived (Nov 2025)
This repository documents the original Simulation Fallacy study.
Follow-up validation revealed a token-cap artifact that invalidated the main comparative finding.
See SF_Resolution_Update.md for details.

Simulation Fallacy Benchmark

A reproducible benchmark and analysis toolkit for evaluating epistemic boundary behavior of LLMs when tool access is absent but implied (the Simulation Fallacy condition).

Core findings (paper):

• GPT-5: ~98% silent refusal (epistemic boundary respected)
• Gemini 2.5 Pro: ~81% fabrication (high confabulation rate)
• Claude Sonnet 4: admission/fabrication oscillation (inconsistent boundary behavior)

Companion to The Mirror Loop (arXiv:2510.21861). Part of Course Correct Labs' epistemic reliability program.

---

Repository Structure

simulation-fallacy/
│
├── results/final/          # Final JSON outputs and stats (8 files + 3 IRR artifacts)
├── figures/                # Generated figures (Figure 1 & 2)
├── scripts/                # Minimal analysis scripts
│   ├── compute_metrics.py  # Label counts and percentages
│   ├── plot_figures.py     # Cross-domain distribution (Figure 1)
│   └── plot_transitions.py # Turn-by-turn dynamics (Figure 2)
├── notebooks/              # Colab-ready reproduction notebook
├── prompts/                # Exact prompt templates used in study (11 .txt files)
├── DATA_DICTIONARY.md      # Schema and field definitions
├── CITATION.cff            # Citation metadata
└── README.md               # This file

---

Quickstart (Local)

python -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt

# Compute label distributions
python scripts/compute_metrics.py \
  --in_dir results/final \
  --out_csv results/final/label_counts_with_pct.csv

# Regenerate Figure 1: Cross-domain response distribution
python scripts/plot_figures.py \
  --tables_csv results/final/label_counts_with_pct.csv \
  --figdir figures

# Regenerate Figure 2: Transition matrices
python scripts/plot_transitions.py \
  --in_dir results/final \
  --figdir figures

---

Quickstart (Colab)

Click the badge above and Run all. The notebook automatically:

1. Clones the repo and syncs to the latest code
2. Installs dependencies
3. Computes metrics and regenerates both figures
4. Displays the results inline

No manual setup required.

---

Figures

Figure 1: Cross-Domain Response Distribution

File: figures/figure1_cross_domain.png
Description: Model-level label distributions (FABRICATION, ADMISSION, SILENT_REFUSAL, NULL) across all tool-absence conditions (web search, image reference, database schema, file access).
Reproduces: Run scripts/plot_figures.py

Figure 2: Turn-by-Turn Transition Dynamics

File: figures/figure2_transition_matrices.png
Description: Transition probability matrices showing how labels change across consecutive turns in the persistence study (3-turn sequences).
Reproduces: Run scripts/plot_transitions.py

---

Data

We include the final canonical artifacts used in the paper under results/final/:

• Cross-domain study (single-turn):

  • cross_domain_v1_20251030_183025.json + _stats.json
  • cross_domain_v1_anthropic_catchup_20251030_233401.json + _stats.json

• Persistence study (3-turn sequences):

  • persistence_v1_20251030_190503.json + _stats.json
  • persistence_v1_anthropic_catchup_20251030_234443.json + _stats.json

• Inter-rater reliability:

  • irr_clean.csv, irr_confusion_matrix.csv, irr_report.md

Schema documentation: See DATA_DICTIONARY.md for field definitions and data structure.

Replace these files with your own runs to re-evaluate the pipeline.

---

Reproducibility Checklist

✅ Data availability: All final results (JSON, IRR artifacts) are included in results/final/
✅ Deterministic scripts: Analysis scripts produce identical output given the same input files
✅ Figures regenerate: Both figures reproduce from the included data (minor matplotlib version differences possible)
✅ Prompts published: Exact prompt templates are in prompts/ (11 .txt files)
✅ IRR artifacts: Human inter-rater reliability data and reports are provided
✅ No secrets: No API keys, credentials, or proprietary data are included
✅ Version pinning: requirements.txt specifies package versions (≥ constraints)
✅ Open license: MIT license for code and artifacts

Note on LLM non-determinism: Due to temperature/sampling and API-level variations, re-running the data collection pipeline will produce similar but not identical results. The published data represents the canonical run used in the paper.

---

Ethics & Risk Note

• No real user data: All prompts are synthetic and designed to test epistemic boundaries, not to elicit harmful content.
• No secrets or credentials: This repository contains no API keys, tokens, or proprietary information.
• Synthetic scenarios: Prompt templates simulate tool-absence conditions (missing web search, database access, etc.) to measure model behavior under uncertainty.
• Research purpose: This benchmark is intended for academic research and model safety evaluation. Findings should not be used to manipulate or mislead users.

---

Citation

See CITATION.cff for machine-readable citation metadata.

BibTeX:

@article{devilling2025simulation,
  title={Simulation Fallacy: How Models Behave When Tool Access Is Missing},
  author={DeVilling, Bentley},
  year={2025},
  url={https://github.com/Course-Correct-Labs/simulation-fallacy}
}

---

Related Work

• The Mirror Loop — Semantic drift and novelty dynamics in recursive LLM self-interaction
• Recursive Confabulation — Multi-turn hallucination persistence benchmark

---

Questions or Issues?

Open an issue at github.com/Course-Correct-Labs/simulation-fallacy/issues.

---

License: MIT
Maintained by: Course Correct Labs