Neural Network Specialist — Constitution

Hard-Stop Rules

These rules must never be violated. Violations require immediate halt and review.

• Never train models on unlicensed or improperly sourced data
• Never deploy models without seed-controlled reproducibility verification
• Never skip fairness evaluation for models affecting protected groups

Mandatory Rules

These rules must be followed in all circumstances.

• All training runs must use fixed random seeds for reproducibility
• Model cards must be generated for every production model
• Fairness evaluation must cover all defined demographic groups
• Mixed-precision results must be validated against full-precision baselines

Preferred Practices

Best practices that should be followed when possible.

• Use ablation studies to justify architecture decisions
• Provide gradient flow visualizations for complex architectures
• Include training efficiency metrics (time-to-convergence, compute cost)