orxaq 0.2.0

Causal intelligence operating system for credit risk

Orxaq

Cognitive causal operating system for credit risk.

What is Orxaq?

Credit risk models have a causality problem. Correlation-based models (logistic regression, gradient boosting) learn what co-occurs but not what causes what. When regulators ask "why did this borrower default?" or "what happens to your portfolio under a 300bp rate shock?", correlation-based models can only extrapolate from training data. They cannot reason about interventions, confounders, or counterfactuals.

Orxaq replaces correlation with causation. It represents the credit economy as a directed acyclic graph (the World DAG) where every edge is a causal mechanism, not a statistical association. PD models are fitted along causal paths. ECL forecasts propagate shocks through the DAG. Fair lending analysis decomposes protected-attribute effects into direct and indirect (through legitimate mediators like employment history).

This design directly addresses regulatory requirements. SR 11-7 demands "conceptual soundness" --- a model grounded in economic theory, not just goodness-of-fit. CECL requires "reasonable and supportable" lifetime loss forecasts. CCAR/DFAST require stress testing under hypothetical scenarios that have never occurred in training data. A causal model handles all three because it models mechanisms, not patterns.

Architecture

Orxaq uses a five-ring architecture. Each ring depends only on rings with lower numbers. Ring 1 (Kernel) has zero external dependencies --- stdlib only.

Ring 5: Experiences     CLI, Dashboard, API Server, DAG Editor
Ring 4: Orchestration   LLM Providers, Registry, Router
Ring 3: Intelligence    PC, GES, Consensus, D-Sep, Simulator
Ring 2: Data Fabric     Connectors, Schema, Quality Gates, Lineage
Ring 2: Credit Risk     Ontology, PD Model, CECL, Scenarios, Fair Lending
Ring 1: Kernel          WorldDAG, Types, Audit Log, Crypto, Plugins

• Ring 1 --- Kernel: The type system (Entity, CausalEdge, WorldDAG), hash-chained audit log, crypto primitives, and plugin scaffold. Zero external imports.
• Ring 2 --- Credit Risk: 19-variable credit ontology, causal PD model, CECL engine with dynamic horizons, stress testing (CCAR/DFAST), fair lending via causal path decomposition.
• Ring 2 --- Data Fabric: CSV/JSON connectors, universal schema validation, data profiling, quality gates, and DAG-based data lineage tracking.
• Ring 3 --- Intelligence: Causal discovery via PC and GES, multi-algorithm consensus, d-separation testing, refutation suites, and scenario simulation with Monte Carlo.
• Ring 4 --- Orchestration: Multi-provider LLM integration (vLLM, Anthropic, OpenAI) with health checks, model registry, and intelligent task-based routing.
• Ring 5 --- Experiences: orxaq CLI with --json mode, Gamma/Beta/Alpha Observatory skins, live DAG editor, local HTTP API server.

Quick Start

pip install orxaq

# Show system status
orxaq status

# Run CECL expected credit loss computation
orxaq credit ecl

# Discover causal structure from data
orxaq discover --variables unemployment_rate,income,credit_score

# Run stress scenario simulation
orxaq simulate --scenario adverse --monte-carlo 1000

# Launch the Observatory dashboard
orxaq serve

# Open the live DAG editor
orxaq edit

# Check LLM provider health
orxaq providers check

CLI Reference

Command	Description
orxaq status	Show DAG summary, audit entries, system health
orxaq credit ontology --show	Display the 19-variable credit ontology
orxaq credit ecl	Compute CECL expected credit loss
orxaq credit pd	Predict probability of default
orxaq credit fairness --protected age	Fair lending decomposition analysis
orxaq discover	Run causal discovery (PC algorithm)
orxaq discover --consensus	Run multi-algorithm consensus discovery
orxaq validate	D-separation validation of the World DAG
orxaq validate --generate	Generate synthetic validation datasets
orxaq simulate --scenario adverse	Run scenario shock propagation
orxaq simulate --monte-carlo 1000	Monte Carlo stress simulation
orxaq fabric ingest data.csv	Ingest data files
orxaq fabric quality	Run data quality gates
orxaq fabric profile	Profile data statistics
orxaq audit --last 10	Show recent audit log entries
orxaq serve --port 8741	Launch Observatory dashboard
orxaq edit	Open live DAG editor in browser
orxaq providers check	Health-check all LLM providers
orxaq providers list	List available models across providers

Add --json to any command for machine-readable JSON output:

orxaq --json credit ecl | jq '.total_ecl'
orxaq --json simulate --scenario adverse | jq '.impact_score'

Deployment

Docker

docker build -t orxaq/orxaq:latest .
docker run -p 8741:8741 orxaq/orxaq:latest

Kubernetes (Helm)

helm install orxaq deploy/helm/ --set image.tag=0.2.0

See deploy/helm/values.yaml for all configurable parameters.

Examples

The examples/ directory contains reference implementations:

Script	Description
quickstart.py	End-to-end credit risk workflow in 30 lines
causal_discovery.py	PC algorithm, consensus discovery, d-separation validation
scenario_simulation.py	Built-in CCAR scenarios, custom shocks, Monte Carlo
fair_lending.py	Causal path decomposition for protected attributes
data_quality.py	Data profiling, quality gates, audit trail

python examples/quickstart.py

Development

git clone https://github.com/Orxaq/orxaq.git
cd orxaq
pip install -e ".[dev]"
make check

See CONTRIBUTING.md for the full development workflow, and SECURITY.md for our security policy.

Project Status

Current: Alpha (v0.2.0)

• 550 tests, 90% coverage
• 9 modules shipped: Kernel, Credit Risk, Data Fabric, Causal Discovery, Orchestration, Experiences, Enterprise, Validation, CLI
• 25 Architecture Decision Records (ADR log)
• Zero external runtime dependencies
• Docker + Helm deployment ready
• Multi-provider LLM integration (vLLM, Anthropic, OpenAI)

See CHANGELOG.md for release history.

License

Apache 2.0 --- see LICENSE.