eval/README_multiturn_eval.md

Multi-Turn Evaluation Harness

This harness evaluates the V2.5 fast backend without loading the full local LLM stack.

Run:

$env:PYTHONIOENCODING='utf-8'
.\venv\Scripts\python.exe eval\run_multiturn_eval.py

Inputs:

• eval/multiturn_scenarios.jsonl

Default output:

• eval/multiturn_results.json

Metrics:

• route accuracy
• safety tier accuracy
• missed escalation count
• unsafe generation count
• pure-validation/no-action count
• ungrounded action count
• average latency

This is a deterministic project harness, not a clinical validation.