epalea 1.0.2

Latent Factor Posteriors — evidential probabilistic inference

Epalea — Latent Posterior Factors (LPF)

Research and systems for multi-evidence reasoning, uncertainty quantification, and probabilistic inference.

LPF (Latent Factor Posteriors) encodes evidence items into latent Gaussian posteriors, then decodes them into soft likelihood factors for tractable Sum-Product Network reasoning.

Quick Install

pip install epalea

5-Minute Quickstart

import epalea

model = epalea.load_model("compliance-v1")  # downloads from HF automatically

result = model.infer(entity_id="C0001", mode="both")
print(result.results.spn.prediction)        
print(result.results) 
print(result.uncertainty)        

CLI

epalea generate-data --domain compliance --n-entities 300
epalea train         --config ./user_workspace/configs/train.yaml
epalea index         --config ./user_workspace/configs/index.yaml
epalea train-aggregator --config ./user_workspace/configs/train_aggregator.yaml
epalea infer         --entity-id C0001 --mode both

Pretrained Weights

Model	Domain	Hugging Face
compliance-v1	Compliance scoring	aaaEpalea/epalea-compliance-v1

Weights download automatically on first use — no manual setup needed:

model = epalea.load_model("compliance-v1")  # downloads from HF on first call

Papers

Latent Posterior Factors (LPF) is developed across a structured research program:

Core Framework

I Know What I Don't Know: Latent Posterior Factor Models for Multi-Evidence Probabilistic Reasoning Introduces LPF as a unified probabilistic framework for aggregating multi-source evidence with calibrated uncertainty.

• arXiv: https://arxiv.org/abs/2603.15670
• Zenodo: https://zenodo.org/records/19183861

Theoretical Foundations of Latent Posterior Factors: Formal Guarantees for Multi-Evidence Reasoning Provides formal guarantees, theoretical grounding, and convergence properties of LPF models.

• arXiv: https://arxiv.org/abs/2603.15674
• Zenodo: https://zenodo.org/records/19184458

---

Architecture & Systems Analysis

Dissecting Hybrid Neuro-Symbolic Systems: An Architectural Ablation Study of the Latent Probabilistic Framework Analyzes LPF components through ablation to identify contributions of architecture modules.

• Zenodo: https://zenodo.org/records/19186405

---

Safety & Decision-Making

Safety-Critical Decision Making for Autonomous Machines: A Probabilistic Framework with Uncertainty-Based Human Handoff Proposes uncertainty-aware decision thresholds enabling safe human intervention.

• Zenodo: https://zenodo.org/records/19187781

---

Generalization & Transfer

Cross-Domain Transfer and Calibration Analysis for Latent Probabilistic Frameworks Studies generalization and calibration behavior of LPF across domains.

• Zenodo: https://zenodo.org/records/19201876

---

Reasoning & Verification

Exact Uncertainty Decomposition for Multi-Evidence Fact Verification: A Formal Alternative to LLM-Based Reasoning Introduces exact decomposition of uncertainty for verifiable multi-evidence reasoning.

• Zenodo: https://zenodo.org/records/19202524

---

Robustness

Robustness Under Noise: A Comprehensive Analysis of Latent Factor Posterior Models in Noisy Compliance Environments Evaluates LPF stability and degradation under structured and unstructured noise.

• Zenodo: https://zenodo.org/records/19203043

---

Scientific Applications

Uncertainty-Aware Quantum State Tomography via Latent Posterior Factors: A Multi-Evidence Aggregation Approach Applies LPF to quantum state reconstruction with uncertainty-aware inference.

• Zenodo: https://zenodo.org/records/19203267

---

Scalability

Scalability of Latent Factor Posteriors to Varying Evidence Pool Sizes Examines how LPF scales with increasing number and diversity of evidence sources.

• Zenodo: https://zenodo.org/records/19203493

---

Interpretability & UX

Understanding Provenance: A User Study on Explainability in Probabilistic Multi-Evidence Reasoning Systems Investigates user understanding of LPF explanations and provenance tracking.

• Zenodo: https://zenodo.org/records/19203617

About Epalea

Epalea is a research-driven effort focused on building reliable and interpretable AI systems grounded in principled probabilistic reasoning.

https://epalea.com

License

• Source code — Apache License 2.0
• Pretrained weights — Epalea Model Weight License

Commercial weight licensing: [e-mail]

Citation

@article{epalea2026lpf,
  title  = {Epalea: Latent Factor Posteriors for Evidential Inference},
  author = {Epalea Team},
  year   = {2026},
  url    = {https://epalea.ai}
}

Documentation

Full documentation at epalea.ai/docs

Star the repository to follow progress.