Model inputs
The simulation uses persisted fighter and division features, including adjusted ratings, activity, modeled finishing tendencies, sample coverage, and volatility. It does not invent injuries, camp reports, or qualitative claims that are absent from the source data.
Monte Carlo simulation
Monte Carlo iterations estimate winner, method, and round distributions under the published model snapshot and deterministic seed strategy. A probability is a model estimate, not certainty; sparse histories, inactivity, volatility, and five-round calibration remain important limitations.
Immutable publication
A complete event card is published in one transaction before fight lock. Its model version, input cutoff, seed, coverage, output hashes, and timestamp remain immutable. A changed forecast becomes a linked revision instead of overwriting the old call.
Grading and calibration
Official results append grading revisions. The public record reports denominators alongside accuracy, Brier score, log loss, calibration bins, expected calibration error, and method and round scores. Draws, no contests, voids, and unresolved results remain visible and are excluded only where a valid scoring target does not exist.
Market provenance
Market comparisons use timestamped, provider-labeled winner quotes. Each observation retains its source identifier, capture time, price type, liquidity, freshness, and no-vig probability. A closing comparison is the last complete valid pair at or before the same fight lock; later quotes are never labeled closing.
Corrections and limits
Forecast corrections create a new linked publication, while official-result corrections append a new grade and rebuild derived metrics. Missing or stale evidence is shown rather than filled with invented values. Historical performance cannot guarantee a future result or profit.
