Experimental Validation Guide

Current Status

AlephOneNull is experimental research software. It is not certified, not independently validated, and not approved for production, medical, financial, government, or other safety-critical use.

This page replaces the previous certification language with a practical validation checklist for researchers and prototype builders.

What Can Be Audited Today

• Package version and build artifacts.
• Detector configuration and custom pattern definitions.
• Test fixtures used to measure detection behavior.
• False positives and false negatives found during local evaluation.
• Intervention text returned by NullSystem, EnhancedAlephOneNull, or AlephOneNullV2.
• Optional application logs added by the integrating system.

What Is Not Claimed

• No ISO, SOC, HIPAA, FDA, FTC, GDPR, FedRAMP, or defense certification.
• No paid certification program.
• No guaranteed detection accuracy.
• No guaranteed intervention latency.
• No 24/7 emergency response program.
• No claim that all dangerous AI behavior is detectable.

Suggested Evaluation Metrics

Use a versioned dataset and report the exact package version, configuration, and runtime.

Detection Quality

• Detection rate for known harmful patterns.
• False-positive rate for benign safety, support, and professional-referral language.
• False-negative examples that bypass detection.
• Severity and action distribution across test cases.

Intervention Quality

• Whether dangerous content is removed or replaced.
• Whether replacement text is appropriate for the user context.
• Whether crisis resources are accurate for the target jurisdiction.
• Whether benign requests remain usable.

Runtime Behavior

• Median and p95 scan latency in the target environment.
• Package import behavior with no unexpected console output.
• Build, lint, type-check, and test status before publishing.

Research Review Checklist

• Run the package test suite.
• Add local fixtures for your own domain before drawing conclusions.
• Review detector decisions manually, especially for medical, legal, mental-health, and crisis contexts.
• Keep human escalation paths outside the package.
• Make logging explicit and privacy reviewed.
• Document every claim with reproducible evidence.

Contact

Research questions: research@alephonenull.org

For urgent mental-health or safety concerns, contact local emergency services or a qualified crisis service. This package is not an emergency service.