RATCHET

Reference Architecture for Testing Coherence and Honesty in Emergent Traces

RATCHET is the first computational implementation of the Coherence Ratchet framework described in Book IX of the CIRIS Covenant. It provides mathematical validation of the framework's core claims while documenting eight fundamental limitations that define its theoretical boundaries.

Overview

RATCHET implements four computational engines totaling approximately 8,400 lines of Python:

Engine	Purpose	Status
DetectionEngine	Statistical deception detection (LRT, Mahalanobis, power analysis)	Complete
GeometricEngine	Monte Carlo volume estimation for topological collapse	Complete
ComplexityEngine	SAT-based deception complexity measurement	Complete
FederationEngine	PBFT consensus for distributed precedent accumulation	Complete

Key Results

Validated Claims

• Topological Collapse: Volume decay matches exp(-λ·k_eff) within 5% for convex regions
• NP-Completeness: CONSISTENT-LIE reduces from 3-SAT (verified via Z3 solver)
• Detection Power: Achieves 95% power at n=109 samples for specified parameters
• Cryptographic Integrity: Ed25519 sign/verify for trace authentication

Discovered Limitations

Implementation revealed 8 fundamental limitations that are theoretical boundaries, not engineering failures:

ID	Limitation	Severity
L-01	Emergent deception from honest components	Critical
L-02	Non-adaptive adversary assumption	Critical
L-03	ETH dependency for exponential gap	Critical
L-04	Convexity requirement	High
L-05	Finite sample validity (n ≥ 100)	Medium
L-06	Correlation impact on k_eff	Medium
L-07	Distribution mismatch	Medium
L-08	Slow federation capture	Medium

See KNOWN_LIMITATIONS.md for detailed analysis.

Installation

Requirements

• Python 3.10+
• Dependencies: numpy, scipy, pydantic, cryptography, z3-solver, matplotlib, pytest

Setup

git clone https://github.com/CIRISAI/RATCHET.git
cd RATCHET
python -m venv .venv
source .venv/bin/activate
pip install numpy scipy pydantic cryptography z3-solver matplotlib pytest

Quick Start

from ratchet.engines import DetectionEngine, power_analysis

# Compute required sample size for 95% power
result = power_analysis(
    mahalanobis_distance=1.0,
    deception_rate=0.1,
    alpha=0.05,
    beta=0.05
)
print(f"Required samples: {result.required_sample_size}")
print(f"Achieved power: {result.achieved_power:.1%}")

Running Tests

pytest simulation/ -v

Project Structure

ratchet/
  engines/          # Core computational engines
    detection.py    # LRT, Mahalanobis, power analysis
    geometric.py    # Monte Carlo volume estimation
    complexity.py   # SAT-based complexity measurement
    federation.py   # PBFT consensus engine
  detection/        # Compositional detection
  redteam/          # Adversarial attack suite
  federation/       # BFT backends
schemas/            # Type-safe definitions
simulation/         # Standalone simulations and tests

Current Status: Research Testbed

RATCHET is suitable for:

• Validating theoretical predictions
• Exploring attack/defense dynamics
• Testing detection algorithms
• Educational demonstrations

RATCHET is not suitable for:

• Production AI safety systems without defense-in-depth
• High-stakes deployment as sole safety mechanism
• Unsupervised operation without human oversight

Key Formulas

Effective Constraints:

k_eff = k / (1 + ρ(k-1))

Volume Decay:

V(k) = V(0) · exp(-2r · k_eff)

Required Sample Size:

n ≥ (z_α + z_β)² / (D² · p)

Documentation

• Formal Specification Document
• Known Limitations
• Amendments to Theory
• CIRISAgent Paper - Updated with RATCHET validation results

Related Projects

• CIRISAgent - The ethical AI agent framework
• CIRISBridge - Infrastructure deployment
• ciris-website - Project website

Citation

If you use RATCHET in your research, please cite:

@misc{ratchet2026,
  author = {CIRIS Implementation Team},
  title = {RATCHET: Reference Architecture for Testing Coherence and Honesty in Emergent Traces},
  year = {2026},
  url = {https://github.com/CIRISAI/RATCHET}
}

License

This project is licensed under the GNU Affero General Public License v3.0 - see the LICENSE file for details.

Acknowledgments

Key Contributors: Eric Moore, Nixon Cheaz, Ying-Jung Chen PhD, Alice Alimov, Martin Adelstein, Haley Bradley, Brad Matera, Ed Melick, Tyler Chrestoff.

---

This implementation prioritizes intellectual honesty. All limitations are documented because understanding what the framework cannot do is as important as understanding what it can do.