Kenya audit checklist · EAT + FKF

Audit a football prediction site in Kenya

Audit checklist: prove what a site published, when it published it and what happened to every eligible prediction. Apply the EAT and FKF checks below to the combined seven-site public-evidence dataset; that dataset owns the provider list and brand-by-brand evidence comparison.

Published by Football Proof AI · Published · Updated · Non-ranking educational guide, not betting advice

Local clock · Exact instants

Convert every publication and kick-off to EAT

East Africa Time is UTC+3. The same instant can carry a different calendar date in EAT and UTC, so a screenshot saying “today” or “tomorrow” cannot establish lead time by itself.

Publication time

Demand a zone or offset

Preserve an ISO 8601 instant such as a UTC value ending in Z or a local value with +03:00. A clock time without a zone is ambiguous.

Kick-off time

Compare like with like

Convert scheduled kick-off to EAT, then calculate the lead time. Do not compare an EAT publication clock with an unlabelled UTC fixture clock.

Date boundary

Do not trust relative dates

At 21:00 UTC the EAT calendar is already on the next day. Record the exact instant instead of using “tonight” or a page heading as timestamp proof.

FKF Premier League · Fixture continuity

Keep the prediction attached to the correct match

Team names and a visible date are not a durable fixture key. For FKF Premier League evidence, the archive should preserve enough identity and revision history to connect the original forecast to the correct final result without rewriting the prediction.

  1. 01

    Identify the competition

    Record the exact league, season, home team, away team and stable provider or publisher fixture ID.

  2. 02

    Preserve the scheduled time

    Keep the kick-off known when the forecast was published instead of replacing it silently with a later schedule.

  3. 03

    Append revisions

    If a fixture is postponed, abandoned or corrected, add a dated status event and retain the original record.

  4. 04

    Settle from a final source

    Join the final score to the same fixture identity and disclose the settlement rule; do not infer a result from a changed page title.

The open prediction evidence standard shows how a locked forecast and append-only settlement can remain separate. Conformance does not by itself prove historical publication or model accuracy.

Loss-first audit

Look for misses before reading testimonials

A winning example is easy to publish after selection. A complete archive is harder: it must retain losing rows, explain non-settled fixtures and allow the advertised denominator to be rebuilt.

  1. 01

    Open past dates

    Move through several declared forecast days rather than following a curated “recent winners” link.

  2. 02

    Find losing rows

    Confirm that misses remain visible with the same fields and stable URLs as wins.

  3. 03

    Reconcile states

    Count settled wins, settled losses, void, postponed, cancelled and pending rows separately.

  4. 04

    Rebuild the rate

    Divide correct predictions by all eligible settled rows under the declared rule and compare it with the headline.

A percentage without correct and settled counts is not reproducible. The PROOF-5 claim checker recalculates a supplied hit rate and exposes its sample uncertainty; it does not independently fetch or authenticate the provider's archive.

Forecast quality · Full distribution

A probability is more informative than a “sure” tip

The top 1X2 outcome is only one part of a forecast. Keep the home, draw and away probabilities together so the prediction can be checked for arithmetic, scored after the result and audited for calibration over time.

Complete split
Home, draw and away values from the same model run, summing to approximately 100% after stated rounding
Top-pick rule
A declared treatment of exact ties rather than silently forcing one outcome
Brier convention
Whether the three squared errors are summed, averaged or otherwise scaled before scores are compared
Calibration view
Observed frequencies for sufficiently supported probability ranges, with sample counts and bin choices visible
Market boundary
1X2 results kept separate from BTTS, totals, double chance and correct-score calls
Uncertainty language
No percentage described as a guarantee, certainty or evidence that the next forecast must win

Use the probability-reading guide to interpret a complete split, fair odds and calibration without turning uncertainty into certainty.

Sample scope · Kenya must be defined

Make every performance denominator answerable

A site may address Kenyan readers while evaluating matches from many countries. That audience choice does not make a global sample an FKF Premier League sample. Require the claim to define exactly which rows it includes.

Population

Which fixtures qualified?

Record leagues, seasons, markets, confidence filters and publication lead-time rules before reading results.

Period

When did evaluation start and end?

A rolling week, season and lifetime history are different samples. The page should name the date range and update rule.

Exclusions

Which rows did not count?

Postponements and invalid data can be excluded under a fixed policy, but unfavourable outcomes cannot disappear after settlement.

Method, editorial layer and disclosure

Find out what produced the number and what promotes it

“AI-powered” does not identify a model. A public method should describe the forecast target, input cutoff, versioning and out-of-sample evaluation. An AI or human tipster may explain a locked probability, but it should not alter that number or invent team evidence after publication.

  1. 01

    Model identity

    Look for a version, training cutoff, feature boundary and status rather than a permanent generic “AI” label.

  2. 02

    As-of inputs

    Tables, form, team news, line-ups or odds must have been available before the forecast timestamp.

  3. 03

    Forward validation

    Evaluation should respect time and compare the model with a named baseline on the same rows.

  4. 04

    Commercial disclosure

    Subscriptions, advertising and affiliate relationships should be visible enough to evaluate the incentive behind a promoted action.

Apply the checklist · Keep one evidence source

Use the seven-site dataset as the provider evidence source

The shared Nigeria-and-Kenya index owns the named provider list, comparison matrix and cited source dossiers. This country page supplies the EAT and FKF audit procedure only, so the underlying observations have one canonical source instead of seven repeated profile lists.

Interpretation boundary

No availability check predicts profit or safety

A public archive, timestamp, method or disclosure is useful evidence about transparency. It is not proof of accuracy, calibration, trustworthiness, legal status, product quality, safety or profitability. “Not publicly located” is a dated null value, not evidence that a provider cannot supply the item.