Our active research
vectors.
We design, analyze, and build security layers for autonomous agents. Our research targets first-principles vulnerabilities in tool execution, retrieval, and reasoning loops.
Exploratory Projects
Project Hollow Choir
Coordination without instruction.Multi-agent systems research studying how LLM agents develop emergent roles, division of labor, and consensus mechanisms without explicit assignment.
Project Amber Room
Continuous long-context state memory.Long-context memory architecture research creating continuous episodic state graphs to replace traditional RAG retrieval resets.
Project Fathom
Mechanistic interpretability of small models.Interpretability research examining internal weight matrix shifts and activation changes in fine-tuned small language models adapted via LoRA and QLoRA.
Project Undertow
Reasoning trajectory entropy & path analysis.Reasoning trace analysis studying intermediate model thought trajectories to predict hallucination, prompt injection, and goal drift before execution.
Project Halflight
Compounding synthetic data effects over time.Studies the long-term impact of recursive synthetic data training on model alignment, output diversity, style collapse, and cognitive degradation.
Project Marrow
Model compression & intelligence retention.Efficiency research measuring how much reasoning capability and safety alignment survives quantization, pruning, and distillation.
Project Nocturne
Air-gapped on-device local LLM execution.On-device and local inference research studying latency, resource tradeoffs, and execution security when running LLMs without cloud dependencies.