Model card

World Cup 2026 contextual model card

Kickoff Lens separates public match facts from model context. The model is designed to explain why a fixture may tilt one way, not to guarantee a result.

Current version

wc-analytics-v2.1-agentic-202606181200 · generated from public schedule, team, squad, result, and stable venue-profile data.

Current coverage

LayerCurrent coverageWhy it matters
Match model context72/104 matchesShows whether a fixture has explainable model inputs beyond a raw percentage.
Factor explanations72/104 matchesLists the concrete drivers behind the estimate: team quality, attack-defense fit, venue, travel, form, and pressure.
Venue and weather layer104/104 matches · 16 venue profilesUses climate, altitude, roof, travel, and kickoff context before more volatile live weather updates are available.
Squad data1248 listed playersLets the model separate team depth and key-player influence from broad ranking strength.
Agent-readable summaries0/104 matchesTurns scoring inputs into reader-facing explanations and scenario notes.
Finished-match calibration24 match sample · direction 58.3% · Brier 0.2Tracks whether the model is drifting and where it misses, instead of hiding uncertainty.
Source-backed late signals0 signalsReserved for reliable public updates such as squad news, result corrections, and late context.

Inputs used

Two-system operating model

The data updater and the analysis Agent are separate layers. The updater handles schedule, score, squad, source freshness, and post-match refresh windows. The analysis layer reads that package and produces context indexes, factor drivers, confidence notes, and recap language. If the updater is rate-limited, the analysis layer keeps using the latest verified package instead of inventing facts.

How the Agent scores a fixture

  1. Start from team baseline: FIFA ranking context, Kickoff Lens radar, squad depth, core-player impact, and current tournament record.
  2. Adjust for matchup shape: attack versus defense, control versus transition risk, execution quality, style volatility, and group-stage pressure.
  3. Apply environment context: venue climate, altitude, roof profile, travel/rest load, kickoff timing, and likely adaptation risk.
  4. Apply source-backed late signals only when reliable public sources expose them. Missing injury, suspension, or lineup data is shown as uncertainty, not guessed.
  5. After full time, compare the estimate against the actual result and keep a rolling calibration log.

Accuracy operating loop

The Agent is tuned as a reading system, not a picks service. It should become more useful by explaining uncertainty better, adding verified late context faster, and learning where the model was overconfident after finished matches.

SignalCurrent behaviorAccuracy rule
Team strengthRanking, radar, squad depth, key-player impact, and tournament record.Strong teams can still be dampened when context, travel, or matchup style opens the balance lane.
Venue and weatherStable climate, altitude, roof, heat, humidity, and travel/rest proxies.Environment can move close matches, but it should not overpower a clear quality gap without verified late weather.
Lineup/newsReserved for source-backed public signals.If injury, suspension, or expected XI data is missing, confidence drops instead of inventing an adjustment.
CalibrationFinished results feed a rolling Brier and direction check.When the model becomes overconfident, stronger lanes are dampened and balance lanes stay visible.
Post-match learningScores and event timelines are compared to pre-match factors.Wrong calls should update the explanation layer so future previews show why uncertainty mattered.

Confidence rules

How the Worker keeps it current

The Cloudflare Worker wakes on a cron schedule, checks the next post-match windows, overlays verified public corrections when upstream caches lag, rebuilds contextual model reads, and keeps Pages functions reading the live package first with static data as fallback.

Known limits

What a restriction changes

API or Agent limits affect freshness, not the scoring framework itself. When live calls are limited, Kickoff Lens uses the most recent verified data package, marks stale or uncertain states, and waits for the next allowed refresh. That is safer than fabricating weather, injury, or event data. Accuracy improves when reliable late signals are available, but transparency improves by showing what is known and what is missing.

Next model upgrades

Open fixture model reads Open source policy