We build the box a model's control code runs in

What we are doing

We are building a machine-checked bound between language models and autonomous machinery. Model-authored control logic runs on separate, deliberately small hardware, and a fixed monitor separately decides whether each proposed command is permitted.

The trusted core is written and checked to fifteen named properties. On the selected ECP5 board, soft-core execution, loader rejection, a 10,000-run loader/VM burn-in, memory-protection checks, seventeen of seventeen jail-break probes and 12, 24, 48 and 96 MHz workload timing captures have passed. In a PC-assisted rig, a program in the board's interpreter moved a physical SO-101 elbow while an out-of-envelope program was refused, and the arm held when the gate withheld its reply. Product I/O, durable recording and fabric-held monitor state are the next milestones.

The architecture, evidence and research are public. Read the research record →

Who is doing it

Oleg Sidorkin leads Endstop. His background combines a PhD in systematic analysis and control with four years as chief architect of IBM's AI incubator, where he built a conversational platform that carried revenue.

Funding would add hardware and firmware engineering, embedded software testing and machinery safety expertise. Candidates have been identified, and preliminary hiring terms have been negotiated. Safety consulting could provide interim capacity while recruiting.

Two companies before this one reached an exit: Applied Intelligence Solutions, acquired through DemandTec into IBM, and GoalProfit, acquired by Digital Wave Technology. Of four granted patents, one uses the same structural pattern as Endstop's capability table: US 11,842,290 binds language-generated structure to a fixed, typed action vocabulary. The patent is part of the founder's prior record. Separately, Endstop has electronic acknowledgements for five provisional application submissions and matching payment receipts dated 22 September 2026. Official filing receipts and dates remain to be reconciled; no Endstop patent grant is claimed. See the IP and diligence scope.

Endstop has not taken a product through certification. No product here has gone through a notified body, and no IEC 61508 or ISO 13849 engagement sits behind the design. The conformity path is written down and staged, which is a different thing from having walked it. On a machinery-safety product, that gap matters. We state it here so it can be assessed early. A design partner with real hardware moves validation from our bench to an application.

Design partners

We are looking for a small number of teams with real hardware and a real reason to put a model into a control loop. A useful partner generally has:

• A platform with no safety-rated envelope, whose commanded motion we can sit in front of: a research rig, a custom actuator setup, a fixed-base arm, or one bounded subsystem of a larger machine such as the manipulator on a quadruped or a mobile base.
• A model that makes a substantive decision inside the control loop.
• An engineer who will argue with us. Disagreement from people who build machines has improved the design more than agreement would have.

You get early access to the hardware, direct influence over the policy vocabulary and interface, and the threat-model and safety-case work that has to happen anyway. We get contact with an application before the design freezes, and we learn which of your workloads need contained generated logic and which need only the supervisor.

Advisory

We also take on analytical work: threat modelling for a machine controlled by a model, safety-case structure for autonomous machinery, and architecture review of an existing containment design.

This work stands on its own. It commits you to nothing on the hardware.

Contact

Email us directly. It reaches the person doing the work.