Independent AI safety researcher and systems architect building evidence-gated methods for multi-agent systems.

I build AI systems that have to show their work before their outputs can authorize action.

My current work asks a practical question: when several AI agents agree, what would justify treating that agreement as evidence rather than repetition? I build small, inspectable tools for provenance, verifier independence, adversarial testing, and fail-closed decisions.

A standard-library-only Python implementation of semantic-entropy screening for answer stability.

• Runs locally against OpenAI-compatible endpoints.
• Includes 45 deterministic self-checks.
• Treats stability as a screening signal, never as truth.
• Publishes the preregistration and negative result from an extension that failed its own criteria.

That last point matters. I kept the simpler version instead of tuning the experiment until it passed.

A small standard-library wrapper that preserves what a local command actually returned.

• Captures exit status, elapsed time, byte counts, and SHA-256 hashes.
• Keeps raw command arguments out of the receipt unless explicitly requested.
• Refuses to overwrite existing output paths.
• States its limits mechanically: unsigned, not independently validated, and not proof of correctness.

• Multi-agent influence, shared context, and shared-evidence dependence
• Verifier independence and authority boundaries
• Claim provenance, reproducible receipts, and negative controls
• Context continuity without silently promoting stale state
• Harness changes that must outperform boring baselines and survive removal tests

A fluent answer is not a verified answer. A clean test is not scientific validation. A failed experiment is useful when its falsifier, inputs, and limits remain visible.

My work therefore follows a short loop:

1. State what would change the decision.
2. Pre-register the expected effect and failure condition.
3. Compare against a simple baseline.
4. Preserve the raw result and the negative case.
5. Promote nothing from self-certification alone.

I founded Difference Theory after a career in enterprise systems, networks, security, migrations, and recovery. Earlier public work included digital publishing and information propagation at scale. I was a founding co-host of The Drill Down and returned for its 500th episode.

Difference Theory · LinkedIn · semantic-entropy

I am open to technical review, adversarial probes, and research conversations about dependable multi-agent systems.