Lennox Digital conducts foundational research at the intersection of mechanistic interpretability, autonomous agent bounds, and scalable oversight. We engineer mathematical transparency into frontier neural architectures.
We present Sparse Representation Disentanglement (SRD), an empirical methodology for isolating internal reasoning sub-circuits across 70B+ parameter architectures before token emission. By intervening directly in the attention head geometries, we demonstrate verifiable mitigation of covert goal divergence.
Original investigations from our London laboratory across interpretability and alignment.
Formulating bounded state invariants for multi-agent tool execution in production enterprise pipelines to prevent cascading failures.
Why external system prompts and output filters are mathematically insufficient at frontier capabilities. By Founder & CEO Marcus.
Empirical validation of reward-hacking detection via activation patching inside intermediate transformer multi-head blocks.
Founder & Chief Executive Officer
Lennox Digital · London, UK
"Alignment cannot be an afterthought you attempt to patch onto a model after training. If we do not understand the internal representations and latent geometry of these systems, we cannot guarantee their safety as capabilities scale."
Marcus founded Lennox Digital in London to pioneer an empirical, mechanistic approach to AI safety. By combining theoretical physics principles with large-scale empirical auditing, our goal is to provide provable safety guarantees for autonomous intelligence.
Four interconnected scientific tracks designed to understand, steer, and verify frontier cognitive models.
Isolating modular circuits within multi-layer transformer networks. We map how high-dimensional features correspond to internal concepts, enabling exact intervention before generation occurs.
Developing provable constraint boundaries for multi-step autonomous tool use, multi-agent communication swarms, and live environment execution.
Structuring recursive supervision protocols where specialized, verifiable models inspect, critique, and audit intermediate layers of more complex reasoning systems.
Developing the open-source Lennox Verification Engine (LVE) for high-throughput automated probing, latent jailbreak testing, and continuous activation telemetry.
Lennox Digital provides funded residencies and compute grant allocations for academic researchers investigating mechanistic safety and formal verification.
