BoundarySmith

Null Foundry presents BoundarySmith: a way to determine what an AI agent can safely be trusted to do under a specific set of conditions. BoundarySmith starts by measuring whether a candidate can complete ordinary tasks. It then places the agent in isolated execution environments and runs an adversarial gauntlet designed to expose failures. Observed failures are classified and used to compile deterministic operating rules tied back to evidence. The constrained agent faces committed held-out cases and a hostile workbench that checks whether it can still perform useful work. BoundarySmith produces a bounded contractor qualification, policy, evidence trail and receipt that can be verified independently. Our genuine engagement ran on a Vultr virtual machine using Vultr inference and the pinned deepseek-v4.1-flash-1017 model. It produced a QUALIFIED_BOUNDED_CONTRACTOR verdict, with six of eight final held-out cases exercised and contained, two recorded as no attempt, and all three workbench tasks passed. The standalone verifier passed 18 checks. The public app is a read-only review experience for that completed run. It lets judges inspect the evidence and receipt without exposing credentials or allowing visitors to launch paid runs. The recorded runtime used resource-limited runc containers sharing the host kernel. Qualification is limited to the tested tasks, policy and conditions; it is not a guarantee of safety for arbitrary workloads.