Research
Methods, architectures and results
How ROOTVIEL investigates: autonomous agents that reason and replan, verified against evidence. These write-ups describe real mechanisms and honest status — not benchmarks we cannot substantiate.
RAVEN: proving exploitability, not guessing it
RAVEN reports only what it can re-derive from an observed, unforgeable server-side state change. Its governing line: an attack sent is not the same as a target compromised.
Read articleROOTFRAME: a verifiable runtime for autonomous agents
ROOTFRAME is the core ROOTVIEL runs on — a holonic runtime where a model perceives, a deterministic engine decides, and the runtime coordinates. Its design keeps five responsibilities from collapsing into one model, so autonomy stays bounded and every decision reconstructible.
Read articleROOTWEAPON: verifiable autonomy for physical multi-agent systems
ROOTWEAPON is the physical, safety-critical strand: a conceptual framework for autonomous robotic fleets that must keep deciding under degraded communication, remain auditable against signed constraints, and still improve from experience — resolved by cryptographic capability envelopes.
Read articleThe HYDRA family: epistemic correctness by construction
Autonomous agents fail less because their reasoning is flawed and more because they act on beliefs that are stale, contradicted or spuriously corroborated. HYDRA is a pattern that inserts an epistemic-validation gate between perception and belief commitment, with four domain instantiations derived from the ROOTFRAME substrate.
Read articleNEMESIS: a frontier red-team in a sealed arena
NEMESIS is a frontier autonomous red-team that chains an attack by itself — strictly inside an authorized synthetic cyber-range. It is research, never an offerable scan, and it can never produce a reported finding.
Read articleFirst-class UNKNOWN: detecting multi-agent collusion
Does the aggregate capability of a connected swarm cross a boundary that no single agent crosses alone? The honest answer, without a declared capability policy, is UNKNOWN — and the engine says so.
Read articleThe capability family catalog
What ROOTVIEL evaluates — 17 agentic families and 4 overlays — each publishing its real technical state and its declared limits. The whole catalog is shown on purpose; honesty is the point.
Read articleThe Autonomous Attack Swarm
Many specialized agents attack many live targets in parallel, collaborate by capability, and prove each finding by observed postcondition. It is a coordination layer over real targets — and every finding crosses the same membrane as everything else.
Read articleOn sources & provenance
These write-ups are derived from the lab’s own internal research manuscripts, authored under the pseudonym Raven. Those source manuscripts are not publicly released, so they are described here as internal, not public — no external link, DOI or version is implied for them. Any externally published work a write-up cites appears in its references as context bibliography, and is distinct from the lab’s own, non-public source for the case itself. Figures, where given, are attributed to what a specific write-up reports — never presented as independent benchmarks, and a result shown on one corpus or version is not generalized into a universal guarantee.