I am an independent systems architect and AI safety researcher specializing in empirical benchmark determinism, autonomous agent containment, and runtime cryptographic verification.

My work focuses on building the foundational, zero-fluff infrastructure required for rigorous frontier model evaluation.

eval-inference-engine - Eliminating non-deterministic scoring variance in capability and safety benchmarks via Monte Carlo perturbation testing.

mcp-shield-audit - Detecting runtime tool poisoning, prompt injection, and schema drift in Model Context Protocol (MCP) server chains using Merkle hash trees.

long-horizon-containment-suite - Standardized benchmarks measuring whether autonomous agents can breach sandbox boundaries or hack reward scorers across multi-hour task horizons.

Coming from a background in rigorous systems engineering and compliance validation, I build software with formal state invariants, property-based verification (Hypothesis), and adversarial test suites.

All of my work is 100% open-source, reproducible, and verifiable at https://github.com/j-arndt