08. Testing & Quality

Automated tests

Confirmed: No tests/ directory, no pytest suite, no CI test job.

Quality is enforced by:

• validate_traces.py / grug_score.py gates before SFT
• GSM8K eval.py metrics
• Manual checklists in STEPS.md
• Visualizer eslint / prettier / tsc scripts only

What is strong vs weak

Strong	Weak
Committed iteration reports with plots	No unit tests for scorers/formatters
Resume-safe generate/compress	No CI for Python pipeline
Documented regularization ablation story	Repro depends on external HF + API

Interview Q&A

Q: How do you know compression preserved logic?
A: Automated validators check facts/options/tails, plus downstream GSM8K accuracy — not human grading of every trace.