The rules are written before the result is known.

1. Official forecasts

An official forecast must come from the publication pipeline, use an eligible model version and be stored with its three-way probabilities, publication time and kick-off known at publication. The target is to publish at least 24 hours before kick-off. A row that fails the official lead-time or integrity rules should be excluded or visibly flagged, not backdated or manually promoted.

Editorial preference does not select the winning-looking fixtures. When a board is ordered by confidence, that order comes from the published probability, not a writer's favourite team or result.

2. Model output and explanatory content

Illustrative percentages in educational guides must be labelled as examples. Demo data must remain distinguishable from an official live ledger in page copy, metadata and exports.

3. Accuracy and model claims

• Count every eligible settled forecast, including misses.
• Report sample size and evaluation period beside performance metrics.
• Use walk-forward testing so future match data cannot enter an earlier prediction.
• Publish probability-quality measures such as Brier score and calibration, not hit rate alone.
• Identify demo, candidate, shadow and active production models accurately.
• Do not describe a synthetic or development test as production accuracy.

4. Data uncertainty and unavailable states

Missing line-ups, postponed fixtures, feed outages and corrected final scores are normal operational risks. Copy should distinguish known data from inference. When official data cannot be loaded, the site should preserve an unavailable state rather than silently substitute a sample record.

5. Commercial influence

The current site does not rank bookmakers or accept paid placement in model output. Any future sponsorship, affiliate relationship or paid editorial placement must be clearly disclosed and must not change an official probability, model evaluation or settled result.