Calibration
Checks whether events assigned similar probabilities actually happen at similar rates over a published sample.
Public model record
How FootmeshAI records model performance
FootmeshAI model record explains how public football prediction performance will be tracked, including calibration, Brier score, log loss, coverage, freshness, sample size and model state labels.
The public calibration record is being prepared. This page first documents the reporting policy so research, shadow and candidate models are not confused with stable production performance.
Model record key questions
Before using model records, answer four questions
The model record page explains how performance should be tracked instead of turning one model output into certainty.
What sample produced this record?
Check the sample window, coverage, freshness and language state before reading any model performance.
How does calibration relate to a match read?
Calibration helps judge probability reliability, but it does not replace match facts, changes or the evidence matrix.
What state is the model in?
Research, shadow, candidate, production-ready and paused states must stay separate in public performance reporting.
Where do I go after reading this?
Return to the match center for a fixture, or open My FootmeshAI to continue saved model reads.
Tracked metrics
Future model records will track these signals
Brier score
Measures probability accuracy for outcome forecasts. Lower is better, and the score should be compared against simple baselines.
Log loss
Penalizes confident wrong probabilities and helps detect whether the model is overconfident on difficult matches.
Coverage and freshness
Shows how many eligible matches received model output and whether the output was updated close enough to kickoff.
Model states
State labels matter more than model names
Research
Used for offline validation and not shown as a production prediction.
Shadow
Runs beside the product for comparison while the page keeps clear availability labels.
Candidate
Eligible for limited display after calibration, coverage and QA checks pass.
Production-ready
Shown publicly only when matchflow-java marks the output as available for the page language.
Paused
Hidden or labelled unavailable when data coverage, freshness or quality falls below the gate.
Publication policy
Outputs that fail gates are not packaged as performance
A model record should only include output with a clear sample window, language state, data coverage and production availability label. Research, simulated, stale, low-coverage or language-mismatched output must not be mixed into public performance reporting.
Review protocol
Review the performance policy before using a model signal
The model record is not meant to display a flattering score. It helps users see which sample produced the result, what baseline it was compared with and which production state the output currently carries.
Step 1
Check the sample window
Read performance only with its time range, competition scope, language state, eligible match count and excluded output clearly visible.
Step 2
Compare against baselines
Treat calibration, Brier score and log loss as useful only when they are compared with simple baselines and enough matches.
Step 3
Carry the state label
Keep research, shadow, candidate, production-ready and paused labels attached when model output moves into match pages or saved work.
Next step
Bring the model record back into your workspace
After reading the model record, return to the match route first: choose a fixture, open the AI brief, ask, review changes, coverage, the evidence matrix and the research session, then return to the match workspace before continuing saved analysis, saved matches and reminder queues.
Match route
Choose a fixture, then continue through AI brief, Evidence review, change watchlist, coverage, evidence matrix, research session and match workspace.
Back to match centerMy FootmeshAI
Return to saved matches, reminders, recent reads and the personal AI brief.
Open My FootmeshAIAnalysis library
Continue saved model reads and evidence notes.
Open analysis librarySaved matches
Keep priority fixtures in the watchlist and return when data changes.
Open watchlist