Twenty Milliseconds: Hold the Edge
Twenty Milliseconds is an educational simulation game that makes the human oversight timing problem visceral. Control fusion plasma for six seconds. Then watch a 20ms AI controller do it better. Humans complete 2 out of 24 shots at real time. The controller completes 24 out of 24. The question is not whether to automate — it is where to draw the oversight line.
Every AI governance framework includes the phrase "human in the loop." Most of the people writing those frameworks have never experienced what the loop actually feels like when the system operates faster than their hands can move.
Twenty Milliseconds gives them that experience. Two minutes. No physics background required.
What the game is
Twenty Milliseconds is a browser-based educational simulation. You control two sliders — heating power and plasma shape — trying to hold a simulated fusion plasma below its stability limit for six seconds. Higher pressure scores more points. Cross the limit and the plasma disrupts, ending the shot early.
The physics is simplified: a four-equation toy model, not a full plasma simulation. The experience is real: you are making decisions against a system that changes faster than you can respond.
Three modes, one lesson
Practice (1/10 speed)
Time is slowed by a factor of ten. Your hands can keep up. You complete 24 out of 24 shots. Mean performance: 12.0 βN·s. The system feels manageable.
Real time
Same plasma. Full speed. Your hands cannot keep up. You complete 2 out of 24 shots. Mean performance: 5.0 βN·s. Most shots end before three seconds. The stability limit shifts, the tearing forecast spikes, and by the time you react, the plasma has already disrupted.
Controller (20ms cycle)
Same plasma. An AI controller making adjustments every 20 milliseconds. It completes 24 out of 24 shots. Mean performance: 16.2 βN·s. The pressure rides just below the limit, continuously adjusted. The controller sees the forecast, adjusts the heating, and corrects the shape faster than you can register the change on screen.
The lesson is in the gap: 2/24 versus 24/24. Not because you lack understanding. Because you lack speed.
The timescale ruler
The game includes an interactive timeline that makes the timing problem precise:
- 20 ms — one controller decision cycle
- 10–50 ms — illustrative mode growth (instability develops)
- 100 ms — pressure response time
- 200 ms — tearing forecast horizon
- 250 ms — human response time
Five controller cycles fit inside the 100ms pressure response window. Ten cycles fit within the 200ms forecast horizon. Twelve cycles complete before a human can respond at all. The controller is not faster by a small margin. It operates in a different temporal regime.
Who it is for
AI leaders and programme teams — make the automation conversation concrete. Which decisions should be human? Which should be delegated? The game gives every participant a shared experience to reference when the conversation turns abstract.
Risk and governance teams — pressure-test "human in the loop" as a concept. If the system operates at 20ms and the human responds at 250ms, what does "in the loop" actually mean? The game makes the question practical, not philosophical.
Builders, students, and curious minds — hands-on learning through gameplay. Direct comparison with an AI controller. Connection to the underlying prediction and feedback science.
The research foundation
Twenty Milliseconds is built on published research, not metaphor:
Seo et al., Nature 626 (2024) — "Avoiding fusion plasma tearing instability with deep reinforcement learning." The experimental foundation for RL-based tearing avoidance.
Rothstein et al., arXiv 2511.08818 (2025) — "Enabling Integrated AI Control on DIII-D." The PACMAN architecture: diagnostics → predictions → controllers → actuators.
PPPL Laboratory Report (September 2026) — "Key decisions in milliseconds." Five experiments, ~20ms control cycle, 200ms tearing forecast window.
The game’s four-equation model is a simplification. The timing is real. The 20ms cycle is real. The 200ms forecast horizon is real. The gap between human reaction time and controller cycle time is real.
What the game simplifies
The game is explicit about what it is not:
- Four-equation toy model, not full magnetohydrodynamics
- Scripted stability limit, not computed from equilibrium
- Perfect forecast assumption, not noisy sensor data
- Rule-based controller, not a learned RL policy
- Score = βN·seconds (pressure-time proxy), not fusion power
These simplifications are the point. The game does not teach plasma physics. It teaches the timing problem in human oversight. The physics is the vehicle. The lesson is the gap.
Why this product belongs in this portfolio
Twenty Milliseconds is the product that makes the governance thesis physical.
Every other product in the decision-intelligence category records, classifies, or enforces decisions after they are made. KEEL seals the conviction ledger. The Glass and The Binnacle seal operational calls to instrument state. Plimsoll classifies AI reliance limits.
Twenty Milliseconds asks the question that precedes all of them: at what speed does the human stop being the decision-maker and start being the auditor? Where is the line between oversight and observation?
The answer is not in a framework. It is in the gap between 2/24 and 24/24.
Twenty Milliseconds is live and free to play.
Richard Leclézio
Enterprise Transformation & AI Delivery Leader