How elamurugu10/runtime-interceptor works
This is a research platform for containing hallucination cascades across multi-agent LLM workflows. Its principal workflow runs an agent architecture with controlled faults and optional handoff inspection, then evaluates interceptor-on versus interceptor-off outcomes. The README also documents adaptive calibration, dataset generation, three orchestration topologies, an API/dashboard, and statistical evaluation. The graph preserves these capabilities; sampled excerpts and the source index support only some internal call relationships, so other documented wiring is shown as unverified.
Main entry points: manager.py Experiment Manager · interceptor.py Handoff Interceptor · sequential.py Sequential Workflow · hierarchical.py Hierarchical Workflow
Components 20
User Interfaces
- CLI
app/cli.py - REST API
app/api/routes.py - Web Dashboard
app/dashboard/static/js/dashboard.js
Agent Workflows
- Sequential Workflow
app/orchestration/sequential.py - Hierarchical Workflow
app/orchestration/hierarchical.py - Parallel Consensus
app/orchestration/parallel_consensus.py - Agent Roles
app/agents
Runtime Interception
- Handoff Interceptor
app/interceptor/interceptor.py - Detection Signals
app/interceptor/signals - Adaptive Calibration
app/interceptor/calibration/calibrator.py - Repair Engine
app/interceptor/repair/repair_engine.py
Research and Evaluation
- Experiment Manager
app/experiments/manager.py - Benchmark Datasets
app/datasets/generator.py - Human Annotations
app/datasets/annotation.py - Fault Injection
app/fault_injection/injector.py - Statistical Evaluation
app/evaluation/runner.py
Models and Persistence
- Persistence
app/storage/repository.py - Provider Factory
app/llm/factory.py - LLM Providers
app/llm/base.py
Other components
- Researcher
Connections 27
- Researcher → CLI: runs commands
- Researcher → REST API: uses API
- Researcher → Web Dashboard: uses dashboard
- CLI → Experiment Manager: runs experiments
app/cli.py - REST API → Experiment Manager: starts experiments
- Web Dashboard → REST API: uses API
- Experiment Manager → Benchmark Datasets: generates tasks
app/experiments/manager.py - Experiment Manager → Fault Injection: configures faults
app/experiments/manager.py - Experiment Manager → Sequential Workflow: runs architecture
app/experiments/manager.py - Experiment Manager → Hierarchical Workflow: runs architecture
app/experiments/manager.py - Experiment Manager → Parallel Consensus: runs architecture
app/experiments/manager.py - Experiment Manager → Statistical Evaluation: calculates metrics
app/experiments/manager.py - Experiment Manager → Persistence: persists results
app/experiments/manager.py - Benchmark Datasets → Fault Injection: injects faults
app/datasets/generator.py - Benchmark Datasets → Human Annotations: supports annotation
- Sequential Workflow → Agent Roles: dispatches roles
app/orchestration/sequential.py - Sequential Workflow → Handoff Interceptor: inspects handoffs
app/orchestration/sequential.py - Sequential Workflow → Fault Injection: injects faults
app/orchestration/sequential.py - Hierarchical Workflow → Agent Roles: dispatches roles
app/orchestration/hierarchical.py - Hierarchical Workflow → Handoff Interceptor: inspects handoffs
app/orchestration/hierarchical.py - Parallel Consensus → Agent Roles: dispatches roles
app/orchestration/parallel_consensus.py - Parallel Consensus → Handoff Interceptor: inspects handoffs
app/orchestration/parallel_consensus.py - Handoff Interceptor → Detection Signals: runs checks
app/interceptor/interceptor.py - Handoff Interceptor → Adaptive Calibration: loads thresholds
app/interceptor/interceptor.py - Handoff Interceptor → Repair Engine: repairs output
app/interceptor/interceptor.py - Statistical Evaluation → Adaptive Calibration: optimizes thresholds
- Provider Factory → LLM Providers: selects provider
app/llm/factory.py