Ohms Cricket forecasts every match from first principles, then prices it against the market. No part of the prediction is set by a language model — the LLM only extracts facts; a deterministic rules engine turns those into bounded strength adjustments, each traced back to its source and rule version.
A logistic win model converts the per-format Elo gap (plus venue and toss context) into a head-to-head win probability. There is no draw in T20/ODI, so the two sides' probabilities sum to one.
A negative-binomial run-total model simulates the full combined match total — not just win/loss — so the board can price totals and handicap markets, and surface divergence wherever the market disagrees.
Strength is rated separately per format (T20, ODI, Test). A side can be elite in one format and middling in another, so ratings never bleed across formats.
For each competition we run Monte-Carlo simulations of the remaining fixtures to estimate Top-4, final and champion probabilities — the bars you see on the standings board.
- The LLM never sets a probability or rating — it only extracts facts.
- Match-model separation is absolute: logistic win% + negative-binomial run totals.
- Full scorelines are simulated, not win/loss alone — totals + handicap need the distribution.
- Every adjustment traces signal → source URL → rule version.
- Backtests are as-of-match and no-look-ahead, scored against the 1/2 uniform baseline.