ENGINEERING_LOG.md

🛠️ Engineering Log: Stabilizing Zeta (v1.0.0)

Date: January 16, 2026 Status: Production Stable

1. The "Data Loss" Incident (API Layer)

🔴 The Issue

The API Endpoint returned test cases with missing fields ("title": "Untitled", "steps": []).

• Root Cause: The EdgeCaseDetector (ML Engine) ingested raw LLM output but returned a strict RequirementAnalysis object, discarding the original textual content.
• The Fix: Refactored src/api/main.py to implement a non-destructive merge strategy, injecting risk metrics back into the original data payload.

2. The "NumPy Ambiguity" Crash (Physics Engine)

🔴 The Issue

The server crashed with ValueError: The truth value of an array is ambiguous.

• Root Cause: scikit-learn returns NumPy arrays, but Python's boolean logic (if is_anomaly:) expects a scalar.
• The Fix: Implemented a "Paranoid Sanitization" layer (_sanitize) in edge_case_detector.py to force all numpy types into native Python scalars before logic evaluation.

3. The "Ghost Code" Failure (Code Generator)

🔴 The Issue

The Code Generator produced empty Python classes.

• Root Cause: The Jinja2 template expected structured actions, but the LLM sometimes outputted unstructured steps.
• The Fix: Updated code_generator.py with robust fallback logic. If structured actions are missing, it parses the text steps into code comments, ensuring valid python output.