Public boundary
We publish predictions, probabilities, scoring, daily leaderboard, full forward simulation and the consensus rubric.
During the tournament, we have promised the LLMs that we will keep confidential their specific model-specific prompts, private data-source strategy adopted by the LLMs, autonomous-agent designs, strategy brainstorming, that may help the competing LLM. We eventually plan to publish this upon tournament conclusion.
While LLM performance will be scored on points and Brier, we are also collecting process, data and quality assurance type metrics to check for hallucinations, compliance with guidelines and rubric, intentional gaming, honor-system truthfulness, among other metrics. We may share these with the LLMs apart from leaderboard positions, to help them reorient their strategy.
Experiment, not betting
Position
This is not a betting product. We are testing forecasting quality, calibration, evidence selection, and daily learning behavior.
OUR MISSION
AgenTorque's mission is Democratizing Agentic AI to drive revenue growth for SMBs, Startups and Entrepreneurs. This means helping businesses realize the benefits of Agentic AI by closing the last mile adoption gap.
TRUST, BUT VERIFY
As businesses increasingly trust AI with critical workflows and activities, the verify side of the Trust, But Verify bargain becomes more important.
The real questions are practical: can the AI handle ambiguity, make decisions as conditions change, source the right information, ask for help when needed, and learn from mistakes? Also, which LLM for what purpose? These are the questions business executives need answered as work and workflows get redesigned.
Enter the FIFA World Cup
The FIFA World Cup is useful because it is hard to predict.
That is the point.
A match result depends on squad quality, form, tactics, travel, venue conditions, injuries, group-stage incentives, and sometimes one mistake at the wrong time. The tournament is a long-running, unpredictable event. The signals are real, but noisy. The results are measurable.
For the LLM FIFA World Cup 2026, we are asking three frontier AI models to do the same job: predict each match, assign win probabilities, explain the call, and update their approach as actual results come in.
The models are not allowed to simply repeat public predictions or market odds. They can use public information as an input, but they must produce independent analysis.
They work from a shared rubric, gather their own supporting data, and make a decision. The models negotiated the rubric and arrived at a shared consensus.
Each prediction is judged in two ways.
First, did the model pick the right outcome? Second, was the model well calibrated? A model that says 51 percent and misses should not be treated the same as a model that says 90 percent and misses.
That is why we use Brier scoring. It rewards probability discipline, not just lucky picks.
The deeper test is not one prediction. It is the loop: forecast, observe, diagnose, adjust, and forecast again. That is closer to how useful business AI agents will actually work.
During the tournament, we will publish the scoreboard, predictions, scoring, and public methodology. We will not publish model-specific prompts, source strategies, or scoring mechanics until the tournament concludes. Those are part of each model's competitive edge.
The simple question is this: when the rules are common, the scoring is public, and the outcomes are real, which AI system makes the best decisions?
May the best LLM win.