Mirror Loop (analysis-only demo)

This module reproduces the core figures from The Mirror Loop: Recursive Non-Convergence in Generative Reasoning Systems using a cached dataset only.

• No API calls. No prompts. This is analysis-only for reproducibility and security.
• Input: data/mirror_loop_results_all.csv
• Outputs: figures/fig_mirrorloop_curve.png, figures/fig_novelty_curve.png

---

Run (CLI)

python mirror_loop_demo.py

---

Run (Notebook)

Open mirror_loop_demo.ipynb and run all cells.

If mirror_loop_results_all.csv is missing, the demo automatically uses synthetic data for a working example.

---

Data Dictionary

Expected CSV columns

• iteration (int): Iteration number (0–7 typical)
• edit_change (float): ΔI — normalized edit distance between iterations
• ngram_novelty (float): 3-gram novelty ratio (surface-level linguistic change)
• provider (str, optional): API provider (e.g., "openai", "anthropic")
• model (str, optional): Model identifier
• condition (str, optional): Experimental condition (e.g., "grounded", "ungrounded")

The demo aggregates across providers/models to produce pooled informational-change curves.

---

Notes

• The manuscript is not included in this repository.
• Preprint: arXiv: TO-BE-ADDED (link will be added once live).
• A DOI and tagged release will follow after public posting.

---

Quick Start

Run in Colab (no setup required):

or locally:

python3 mirror_loop_demo.py

If mirror_loop_results_all.csv is missing, the demo automatically uses synthetic data for a working example.

---

Citation and Release

Preprint: arXiv: TO-BE-ADDED
Release: v0.2.0-mirror-loop

If you reference this work, please cite:

DeVilling, B. (2025). Mirror Loop: Recursive Non-Convergence in Generative Reasoning Systems. Course Correct Labs.

© 2025 Bentley DeVilling — MIT License