llmiotsafe/EGPv2

EGPv2

EGPv2 is a workflow-style extension of EGP:

• It does not expose SQ1-SQ5 labels to the model.
• It lets the LLM participate in intermediate modules instead of using only static evidence extraction.
• It separates roles into:
  • Triage: infer latent task profile and choose focus chunks
  • Investigator: build competing normal/anomaly hypotheses
  • Supervisor: check whether the analysis is on-topic and sufficient
  • Verifier: make the final structured decision

Files

• api_client.py: OpenAI-compatible API client, optional API key, supports --no-thinking
• benchmark.py: episode collection and resume helpers
• signals.py: deterministic chunking and structured signal extraction
• prompts.py: role-specific prompts
• pipeline.py: orchestration of the workflow
• run_egpv2.py: CLI runner

Quick test

python EGPv2/run_egpv2.py --preview-only --max-episodes 1 --output-dir results/egpv2_preview

preview-only only materializes prompts and traces. It does not call a model, so UNPARSEABLE in the preview summary is expected.

Local vLLM example

python EGPv2/run_egpv2.py \
  --model Qwen/Qwen3.5-9B \
  --api-base http://localhost:8000/v1 \
  --no_thinking \
  --max-tokens 4096 \
  --max-episodes 5 \
  --workers 2 \
  --output-dir results/qwen35_egpv2_5ep