M15: Multiscale Post-Hoc Protocol for Auditing Geometric Fidelity of Latent Projections

Overview

This repository provides the companion code and data for the paper:

M15: A Multiscale Post-Hoc Protocol for Auditing the Geometric Fidelity of Latent Projections Rubén Abella & José Picón (2026)

M15 is a three-layer post-hoc audit protocol that evaluates whether a 2D projection of a latent space preserves sufficient metric, topological, and local semantic-structural integrity for responsible interpretation.

The Three Layers

Layer	Method	What it detects
Global Metric Fidelity	Mantel Test (Pearson on distance matrices)	Wholesale distance distortions, random projections
Hierarchical Topological Stability	PH0 Profile over MST (Wasserstein-1, MST ratio)	Bridge collapses, skeleton deformations, cluster merges
Local Semantic-Structural Dissonance	Fissure Index (Mahalanobis + community-constrained permutation)	Nodes whose semantic profile diverges from their structural community

Key Results

The combined classifier achieves 95.9% accuracy, 96.1% sensitivity, and 95.0% specificity on an adversarial benchmark of 10 scenarios (1,000 Monte Carlo simulations, N=100 nodes).

Repository Structure

motor15-geometric-audit/
├── code/
│   ├── m15_metrics.py                  # Core M15 audit engine
│   ├── synthetic_data_generator.py     # Adversarial scenario generator (A-H)
│   ├── reproduce_benchmark_results.py  # Full benchmark runner
│   └── anonymize_dataset.py           # Dataset anonymization utility
├── data/
│   ├── benchmark_raw_results_v2.csv    # Raw benchmark results (1,000 simulations)
│   └── dataset_anonymized.csv          # Anonymized empirical case study (N=457)
├── paper/
│   └── TRACE_M15_PAPER.pdf             # Full paper (Spanish)
├── pseudocode/
│   └── pipeline_m15_classification.md  # Ordinal classifier pseudocode
├── LICENSE                             # MIT (code)
├── LICENSE-DATA                        # CC-BY 4.0 (paper + data)
├── CITATION.cff                        # Machine-readable citation
└── requirements.txt                    # Python dependencies

Quick Start

# Clone the repository
git clone https://github.com/joinsnipe/motor15-geometric-audit.git
cd motor15-geometric-audit

# Install dependencies
pip install -r requirements.txt

# Reproduce the benchmark (10 scenarios × 100 simulations)
python code/reproduce_benchmark_results.py

Expected Output

The script outputs raw metric values for each scenario. Expected metric ranges:

Scenario	Mantel r_M	W1_norm	D_HT
A_CLEAN	> 0.95	< 0.50	< 0.01
B1_LOW_NOISE	> 0.90	< 0.55	< 0.02
B2_MEDIUM_NOISE	0.75–0.90	variable	variable
B3_HIGH_NOISE	< 0.75	> 0.65	variable
C_RANDOM	< 0.10	> 1.0	> 0.10
D_BRIDGE_COLLAPSE	variable	> 0.65	variable
E_FALSE_OUTLIER	> 0.95	< 0.55	< 0.02
F_REAL_FISSURE	> 0.95	< 0.55	< 0.02
G_CLUSTER_SWAP	< 0.50	variable	variable
H_SPECTRAL_REWIRE	variable	variable	variable

To apply the ordinal classification (PASS/WARN/ALERT/FAIL) to these metrics, see the paper (Section 4) for a description of the classification protocol.

Not Released

The following components are not released and remain available under commercial license:

• Calibrated classification thresholds (PASS/WARN/ALERT/FAIL decision boundaries)
• Production implementation of TRACE Suite engines (M01–M16)
• Specific LLM prompts and orchestration logic
• Nominal raw dataset with actor identities
• Commercial reporting and visualization layers

Authors

• Rubén Abella — Capriciousecret S.L. / TRACE Research (Conceptualization, Methodology, Formal analysis, Software, Writing)
• José Picón — Neural Ghost (OSINT empirical case study, Data curation, Resources)

Citation

If you use this code or data, please cite:

@software{abella2026m15,
  author       = {Abella, Rubén and Picón, José},
  title        = {M15: A Multiscale Post-Hoc Protocol for Auditing the Geometric Fidelity of Latent Projections},
  year         = 2026,
  publisher    = {Zenodo},
  doi          = {10.5281/zenodo.20700231},
  url          = {https://doi.org/10.5281/zenodo.20700231}
}

License

• Code in code/ is licensed under Apache License 2.0 (see LICENSE).
• Paper in paper/ and data in data/ are licensed under Creative Commons Attribution 4.0 International (CC-BY 4.0) (see LICENSE-DATA).

Both licenses require attribution. For commercial licensing of the non-released production code, contact: ruben@spetrace.com