Different models, different jobs, kept strictly separate
A lot of "AI betting bots" quietly ask a language model to guess a score. BetLLM doesn't — probabilities come from statistical models trained on structured match data, validated the way a quantitative researcher would, not from asking an LLM "who do you think wins?"
Four models, one per stage
A separate model for the draft and for each live checkpoint after it, each trained and validated on its own slice of the data. The exact methodology isn't public — but every prediction they produce is tracked openly here, win or lose.
A separate AI for text
Runs entirely on the operator's own machine, nothing sent to a third party. Only reads and classifies text — it never touches the probability or the betting decision.
NO BET by design
A first-class outcome, not an error state. It shows up most often when one of the two teams simply doesn't have much history yet — just a handful of tracked games — which can make a prediction for that match less reliable.
The dataset behind every signal
Every prediction is grounded in a growing, real dataset — not a handful of cherry-picked games.
Validation rules that can't be quietly skipped
No look-ahead bias
Any information that only becomes available after a map starts is never used to predict that same map — otherwise the result would mean nothing.
Honest validation
Every model is checked against games it never saw during training — including the runs that came back average, not only the good ones.
Traceable, always
Every prediction is checked against what actually happened — the result gets published whether the model was right or wrong.
Live checkpoints, not a draft-stage guess
The model re-evaluates a map at several points as it progresses. The earliest, draft-stage checkpoint is the weakest — there's no live game information yet, only team form and draft composition. Checkpoints taken once real net worth, kills and item timings exist are consistently the ones worth watching — all of it measured against more than 5,400 analyzed maps.