Investigate
Read the system briefs, change the controls and measure the response. Use your 2,000 simulator steps per system to investigate the questions that matter.
Ten black-box simulators. A limited experiment budget. A week to turn careful research into models that can see further ahead.
An individual competition. Research with Gemma and submit your own forecasters.You choose the experiments. Your models explain what follows.
Read the system briefs, change the controls and measure the response. Use your 2,000 simulator steps per system to investigate the questions that matter.
Use Gemma to interpret documents, compare hypotheses and develop code. Get your API key from Google AI Studio; Gemma calls do not spend simulator steps. Fit numerical models to the data you collect.
Upload one ZIP with the systems you’re ready to submit. Each forecaster predicts 4,000 steps from an initial observation and a future action schedule.
Move through the illustrations and try a different control in each.
Change initial protection, then watch infection move through a neighborhood.
Adjust buying activity and watch goods leave the market shelves.
Change the signal split and watch queues form on two crossing routes.
Increase demand to see stored reserves support a neighborhood.
Adjust dispatch speed and follow packages from the warehouse to the store.
Restore habitat and watch the landscape support more wildlife.
Open the outlet and watch water flow from the reservoir downstream.
Raise a campaign’s bid to see it win more placements in a shared audience.
Increase bridge outreach and watch an idea move between two groups.
Adjust staffing and watch patients move from the waiting room into care.
Illustrations only. Controls and behavior are simplified, separate from the competition simulators and data.
episodes × forecast steps
for every system
Every participant studies the same underlying parameters and mechanisms. Evaluation changes the starting observations and intervention schedules, using the same initialization rules.
Hidden episodes cover sustained operation, action order, recovery spacing and combined controls. Your uploaded model forecasts the complete trajectory without further observations or network access.
Read the scoring rulesSeptember 2026 · All times are Toronto local time.
Start researching and submit models for public evaluation.
At 12:00 noon. Upload your final models explicitly; public submissions do not carry over.
At 12:00 noon. Final standings follow at a successful sweep once selected evaluations finish.
The public leaderboard opens September 23 at 9:00 a.m. Toronto time. Subsequent standings are swept every six hours at midnight, 6 a.m., noon, and 6 p.m. Toronto time; unfinished selected evaluations delay publication.
Three accepted uploads per system per day, shared across public and final submissions.
The kit includes the research client, system briefs and a numerical baseline.