Free independent checklist · No signup

AI football prediction claim checker

Direct answer: a credible prediction site should let you prove when each forecast was published, reconstruct every result and evaluate complete probabilities against a time-safe baseline. PROOF-5 separates publication, record, metric, out-of-sample and fair-comparison evidence from unsupported claims.

Canonical publication record

Abstract

A browser-local audit of an AI football prediction accuracy claim using recalculated hit rate, Wilson uncertainty and five public-evidence gates for provenance, completeness, metric definition, out-of-sample validation and fair comparison.

Author and publisher
Football Proof AI
Technical report
proof-5-football-prediction-claim-audit/1.0.0
Published
Last modified
Release status
Current release
Review status
Editorial technical note; not externally peer reviewed
Version history
  1. proof-5-football-prediction-claim-audit/1.0.0 : Initial public release.

Interactive PROOF-5 claim audit

Can this AI football accuracy claim be audited?

Enter the advertised hit-rate sample, then mark only what a normal reader can inspect. Marketing statements remain visible as claims but never become proof on their own.

Runs locally in this browserNo URL is fetched, uploaded or stored.
Define the advertised top-pick hit-rate claim
Recalculated90.0%Wilson 95%: 59.6%–98.2%Claim comparison allows ±0.5 percentage points for stated rounding.
01publication provenancePublic pre-match publication proof

Can a reader establish that this exact forecast existed before kick-off?

What public proof could look like

A stable public record or independent archive showing the exact forecast, publication time, scheduled kick-off and genuine pre-match lead time.

02publication provenanceImmutable original and visible revisions

Can later edits be detected without erasing the initially published probability?

What public proof could look like

A content hash, signed receipt or append-only revision trail that distinguishes the original forecast from later corrections.

03record completenessComplete eligible prediction ledger

Can the denominator be reconstructed without trusting a curated wins page?

What public proof could look like

A browseable or downloadable ledger containing every eligible call, including misses, with stable record identities.

04record completenessDefined sample, window and exclusions

Does every headline percentage state what was counted and what was left out?

What public proof could look like

Settled count, start and end dates, leagues, markets, model versions, exclusions and low-sample warnings.

05outcome metricComplete probability and metric definition

Is the claimed outcome or market unambiguous and reproducible from row-level data?

What public proof could look like

The original full probability vector across every mutually exclusive outcome in the claimed market, plus an explicit definition of what the advertised accuracy metric counts.

06outcome metricFixed and consistent settlement rules

Are results graded by the same rule whether the forecast wins or loses?

What public proof could look like

Pre-declared result-source, postponement and void rules applied consistently, ideally through an automatic settlement pipeline.

07out of sampleModel and data version identity

Can a reader tell which reproducible model state produced the forecast?

What public proof could look like

A model version, feature schema, training cutoff and source snapshot linked to the forecast or evaluation window.

08out of sampleTime-safe out-of-sample evaluation

Were every feature and training outcome available before the prediction being evaluated?

What public proof could look like

Walk-forward or rolling-origin folds with point-in-time features, separate calibration data and explicit leakage checks.

09fair comparisonProper scores, calibration and uncertainty

Does the evaluation measure probability quality rather than only correct top picks?

What public proof could look like

Brier score or log loss, calibration buckets or curves, and uncertainty intervals beside top-pick hit rate.

10fair comparisonRelevant, frozen comparison baselines

Is model performance compared with a hurdle chosen before seeing the final result?

What public proof could look like

The same settled rows scored against a frozen base-rate, common-outcome, prior-model or de-margined market reference.

PROOF-5 contract

Five gates, no invented trust score

Each gate contains two observable checks. A gate is complete only when both checks have public proof. A provider statement without a readable supporting record remains partial, however convincing the marketing sounds.

P · Publication provenance
The exact forecast can be traced to a genuinely pre-match public record.
R · Record completeness
The advertised denominator can be rebuilt from a complete eligible ledger.
O · Outcome and metric definition
The probability target, settlement rule and headline calculation are fixed.
O · Out-of-sample protocol
Model evaluation uses only information available before each prediction.
F · Fair comparison
Probability quality, uncertainty and a frozen baseline are reported together.
Claim only
A useful lead for further checking, but not evidence until the supporting artifact is inspectable
JSON report
Deterministic record of the user's selections, protocol version and explicit non-claims

Result meanings

Audit-ready does not mean independently verified

Contradiction found
The reported percentage does not match the supplied correct and settled counts.
Not auditable yet
At least one PROOF-5 gate is missing or supported only by a provider claim.
Audit-ready
All ten checks have user-declared public proof; the underlying artifacts still need independent verification.

No result is a provider rating, certification or promise that a model is accurate, profitable or trustworthy.

Efficient audit order

Verify one prediction before trusting an aggregate

Start with the smallest falsifiable unit: one exact forecast and its publication trail. Then test whether the advertised performance page can be rebuilt from every eligible record.

  1. 01

    Open one record

    Find the complete probability vector for the declared market, original publication time, kick-off and model identity.

  2. 02

    Trace the archive

    Confirm that misses remain browseable and edits create revisions instead of replacing the original.

  3. 03

    Rebuild the sample

    Match the headline denominator to its dates, leagues, markets, versions and exclusions.

  4. 04

    Recalculate quality

    Use proper probability scores, calibration, uncertainty and a frozen baseline on the same rows.

Evidence boundary

A checklist records observations; it does not authenticate them

This page never opens the entered URL. “Public proof” means the user says they found an inspectable artifact; it does not mean Football Proof AIverified the publisher, archive, hash, pipeline or data source.

  • A matching hash proves byte identity, not historical publication.Use an independent archive, witness or signed publication trail when timing matters.
  • A complete ledger proves transparency, not skill.Recalculate scores and compare the same rows with a relevant frozen baseline.
  • A strong backtest does not guarantee deployment performance.Monitor drift, model changes and future out-of-sample results.
  • No evidence audit proves betting value.Odds, transaction costs, limits and harmful gambling risk are outside this protocol.

Authoritative foundations

Why timing, probability scores and version identity matter

The protocol combines established timestamp, canonicalization, hashing, proper-scoring and time-ordered validation building blocks. These sources do not endorse this checklist or certify any result.

  1. RFC 3339.Date and time on the Internet: timestamps.
  2. RFC 8785.JSON Canonicalization Scheme.
  3. Brier (1950).Verification of forecasts expressed in probability.
  4. TimeSeriesSplit.Time-ordered cross-validation reference.