lazarevka research
We build agents that verify their own work. orca and coral are the first two.
build
the bet: capability lives in the harness, not the model.
orca
vuln research & verification agent, authorization-gated. a finding counts only once it reproduces in an ephemeral microVM — the environment issues the verdict, never the model.
coral
a coding agent as one rust binary — every command runs inside a real os sandbox, the run stops at a budget in cents, and it answers to any openai-compatible endpoint.
research
where the products come from.
- machine intelligence reading what models do at inference time; measuring agents over long horizons ongoing
- cryptography & consensus consensus under realistic network assumptions, practical zk, mechanism design ongoing
- robotics closing the sim-to-real gap for embodied agents, measured on hardware ongoing
- computation verifiable compute, zk-vms, compute markets — proving and pricing computation ongoing
about
one of twenty villages, not a valley.
the research arm of lazarevka. we fund our own work and build when building teaches more than reading. orca is the first thing that survived.
- part of
- lazarevka
- shape
- distributed, small
- funding
- private means, no lps
contact
one inbox.