Terminus Est Lab
State-of-the-art AI lab capability.
Terminus Est AI maintains state-of-the-art AI lab capability. The Lab backs the advice and implementation we deliver. Here is some of its recent work.
Selected work
Selected work.
Jane Street Dormant LLM Puzzle
Jane Street's Dormant LLM Puzzle hid backdoor triggers in large language models that otherwise behaved normally. We built methods for finding those triggers — a detection pipeline of probes, baselines, and cross-model comparison — without knowing the training recipe. That work is how we check whether a model slated for military or first-responder use does what the vendor says, and nothing else.
View our code on GitHub →Kaggle Pokémon TCG AI Battle
The Pokémon Trading Card Game AI Battle is a contest to build agents that play a card game with hidden information. We built an information-set Monte Carlo tree search stack for that setting, so the agent chooses one action without seeing the opponent's hand. Most military and first-responder decisions work the same way: you have to act on an incomplete picture.
View our code on GitHub →OpenAI Parameter Golf
OpenAI's Parameter Golf challenge asked for the strongest language model that still fits in a 16MB artifact. We trained and tested model designs under that cap. The same constraint shows up when a model has to run on your own hardware, at the edge, or without a cloud connection.
View our code on GitHub →MDMP/NIMS and ICS in Lean 4
The Military Decision-Making Process (MDMP) and the National Incident Management System (NIMS), including its Incident Command System (ICS), are the planning and command rules used in military operations and emergency response. We wrote core rules — authority chains, course-of-action screening, span of control, unity of command, and life-safety priority — as machine-checked specifications in Lean 4. That makes the gates checkable, which is the point of verifiable decision support.
View our code on GitHub →Wargame decision agents
Star Fleet Battles and Dominions are two of the most complex wargames made: fog of war, timing constraints, and a large space of possible courses of action. We built course-of-action search and decision agents for both, as a way to rehearse hard decisions under time pressure and an adversary. That research informs how we design decision support for operations and incident response.
View our code on GitHub →
Get in touch
Let's talk.
To discuss AI advisement or implementation for your organization, we would be glad to hear from you.
Get in Touch