BookWriter-A2A (RQ1)

This folder contains evaluation scripts and the reference trajectory configuration for the BookWriter-A2A runs under the RESULTS directory.

Files

evaluate_trajectory.py
- Evaluates multiple trajectory metrics (e.g., exact match, in-order match, any-order match, precision, recall).
- Input: per-session execution_path.md files under each model's test_results/<session>/ directory.
- Output: a CSV summary (and optionally a CSV of failure reasons when enabled).
evaluate_success.py
- Computes per-model success rate based on generated artifacts in each session folder.
- Output: JSON (detailed) and CSV (summary).
analyze_retry_patterns.py
- Analyzes retry patterns and error locations from execution_path.md.
- Output: JSON/CSV summaries.
reference_trajectory.yaml
- Defines the reference (ground-truth) trajectory and evaluation configuration.

Run from this directory:

Trajectory evaluation:
- python3 evaluate_trajectory.py --config reference_trajectory.yaml --output evaluation_results.csv
Trajectory evaluation with failure diagnosis (CSV only):
- python3 evaluate_trajectory.py --config reference_trajectory.yaml --output evaluation_results.csv --diagnose-failures
Success-rate analysis:
- python3 evaluate_success.py
Retry-pattern analysis:
- python3 analyze_retry_patterns.py

These scripts assume the default RESULTS layout used in this repository (see the BASE_DIR / PROJECT_NAME constants in each script).
Markdown report generation is intentionally disabled in the trajectory evaluator; outputs are written as CSV/JSON only.