LLMs are not
a black box.
Eigan is a research house focused on mechanistic interpretability. We believe research on open-source models is critical to understanding LLMs — and to deploying them safely and securely in mission-critical systems.
Spectre
Intercepts every request and response, runs the analysis pipeline, gates on policy, and signs an audit record for the transparency log. Works with open weights through layer hooks, and with closed APIs through behavioural analysis.
Detail page — placeholderConcept probes
Concepts are assembled across a region of the model's depth, not at a single layer, and their direction rotates until it settles. Reading the whole trajectory gives a more accurate, more stable probe than the single-layer default.
Read the research →Appliance
[Placeholder] The on-premises deployment: a hardened appliance that runs inference, analysis and audit inside your own network, with policy and identity attached to the workload.
Detail page — placeholder