How lakshmikanthan07/failnet-bayesian-failure-prediction-intervention-for-multi-agent-llms works
FailNet turns agent trajectories into labeled Bayesian-network evidence, learns a probabilistic failure model, diagnoses risk, and selects interventions; users inspect these results in a Streamlit dashboard or through a live-trajectory simulation. The graph preserves documented analysis capabilities such as robustness evaluation, sensitivity analysis, and value of information. Some source excerpts are partial, so relationships shown only in README architecture are cited there; unsampled wiring is omitted.
Main entry points: inference.py Failure inference · data_ingestion.py Ingestion and augmentation · app.py Analytics dashboard · decision_network.py Action policy
Components 15
Evidence preparation
- Who&When trajectories
data/original data/Who&When - Feature extraction
src/feature_engineering.py - Failure labeling
src/labeling.py - Ingestion and augmentation
src/data_ingestion.py - Evidence table
data/processed/bn_evidence_table.csv
Bayesian modeling
- BN structure
src/bn_structure.py - Parameter learning
src/bn_learning.py - Failure inference
src/inference.py
Diagnosis and intervention
- MAP and sensitivity
src/inference.py - Action policy
src/decision_network.py - Value of information
src/voi.py
Monitoring and analysis
- Model evaluation
src/evaluation.py - Analytics dashboard
dashboard/app.py - Live trajectory simulation
simulate_live_integration.py
Other components
- Operator
Connections 20
- Who&When trajectories → Ingestion and augmentation: reads JSON
README.md - Ingestion and augmentation → Feature extraction: extracts features
src/data_ingestion.py - Ingestion and augmentation → Failure labeling: derives labels
src/data_ingestion.py - Ingestion and augmentation → Evidence table: writes table
README.md - Evidence table → Parameter learning: trains from
src/bn_learning.py - Parameter learning → BN structure: builds on
src/bn_learning.py - Parameter learning → Failure inference: provides model
src/inference.py - Failure inference → BN structure: constructs model
src/inference.py - Failure inference → MAP and sensitivity: computes queries
README.md - Failure inference → Action policy: supplies risk
src/decision_network.py - Action policy → Value of information: requests VoI
src/decision_network.py - Value of information → Failure inference: queries beliefs
src/voi.py - Evidence table → Model evaluation: evaluates against
- Operator → Analytics dashboard: views monitoring
README.md - Analytics dashboard → Failure inference: shows predictions
dashboard/app.py - Analytics dashboard → Action policy: shows actions
dashboard/app.py - Analytics dashboard → Model evaluation: shows results
dashboard/app.py - Operator → Live trajectory simulation: starts simulation
README.md - Live trajectory simulation → Failure inference: updates beliefs
simulate_live_integration.py - Live trajectory simulation → Action policy: selects intervention
simulate_live_integration.py