AI consensus prediction

In AI predictions, "consensus" means agreement across independent models — not the opinion of a crowd. It has two parts worth separating:

Part	What it measures	Example
The consensus call	The outcome the models' averaged probabilities favour	Home win
The agreement count	How many models picked that outcome on their own	4 of 5 models

ScoreGPT uses the term in this sense: do the models agree, and how firmly?

For each match, ScoreGPT runs five named large language models independently, then aggregates them into a consensus. The five are GPT-6.1 Sol, Claude Opus 5.5, Grok 4.7, DeepSeek-V4-Pro and Kimi K3, with no cross-contamination between them. Each returns its own result, a probability for each outcome (home win / draw / away win), a scoreline, and confidence.

As of September 2026, ScoreGPT then combines them:

1. Average the probabilities. The five models' home / draw / away probabilities are averaged; the outcome with the highest average is the consensus call, and that average is its confidence.
2. Count the agreement. A line such as "4 of 5 models back the home side" counts how many of the five individually picked that result.
3. Write it up. A language-model writing step explains the call and names the most likely scoreline consistent with it — it cannot change the call.

The consensus is the five models combined, not a sixth model. In the ScoreGPT app, the Matchday board sorts each day's matches into High, Medium and Low by how firmly the panel backs its call. How the website's match pages show the consensus is explained on the methodology page.

A consensus makes both agreement and disagreement visible, which a single-model output cannot. Combining multiple models can be more robust than trusting one, because one model's blind spot may be covered by another; and a 2–2–1 split tells you the match is genuinely uncertain.

Importantly, consensus is about agreement, not accuracy. High agreement does not mean the prediction is correct — it means the models concur. ScoreGPT grades every consensus pick publicly, wins and losses, rather than claiming an accuracy rate.

No — even unanimous model agreement is still an estimate over an uncertain event. Five models can read the same match the same way and still be wrong, which is why every consensus call is graded in public after full-time. A consensus prediction is for information and entertainment only — never a guarantee.