Vertical scrolling shooter
A released original game. A bounded world with explicit rules, repeatable interactions, and observable outcomes.
Released gameSimulation & Testbeds
A world with rules.
A result we can measure.
Controlled environments let us evaluate planning, autonomy, adversarial behavior, reliability, and agent performance. Games are one useful class of testbed because they provide measurable goals, repeatable scenarios, and clear failure states.
Give the agent a world.
Every action changes the environment. Every observation gives us something to inspect.
ObservePlanActMeasure
Earlier work.
Independent games and experiments, since 2016.
Rhythm-game engine
An unreleased engine and prototype. Timing, input, and state provide a precise foundation for repeatable tests.
Engine and prototypeRoguelike experiments
Experiments in rules, exploration, and decision-making inside a changing environment.
Experimental workNarrative experiments
Interactive software exploring choices, branching state, and the consequences of a sequence of actions.
Experimental workNew questions.
Earlier games can now serve as controlled test environments.
This is a present-day use for historical software. It does not mean that modern AI or security research was performed when those games were originally created.
- Planning, autonomy, and automated gameplay
- Agent behavior and measurable task performance
- Adversarial behavior and security
- Reliability, regression testing, and automated QA
- Telemetry, failure states, and failure analysis
Specific experiments belong here when they have a defined setup, retained evidence, and a result that can be inspected.
Read the engineering notes