OUR FORECASTS, REVIEWED — METHODS
31 August–6 September 2026 · Pilot edition

Dates and source. Kickoffs from 00:00 on 31 August to before 00:00 on 7 September 2026, Europe/Bucharest. That is 30 August 21:00–6 September 21:00 UTC. Existing production records were read on 14 September, shortly after midnight Bucharest time. This pilot reviews an earlier completed week; the evidence was retrieved on 14 September, not captured on the original Monday.

Forecast integrity. We used the existing frozen ledgers. Saved generation and evidence-cutoff timestamps precede kickoff. The freeze field is a logical kickoff stamp, not an independent recording of when the freezing job ran. Sirius includes four scorer versions (2.6.0, 2.6.1, 2.7.1, 2.8.0); Janus uses 2.9.1.

Coverage. The database contains 340 fixtures in this period. Sirius has 190 scored forecasts, 149 without a forecast and one postponed forecast. All-match error is 0.6130; the separate 127-match comparison is 0.6071 versus 0.6047. One legacy row lacks saved home/away IDs; retaining it matches the tracker. Excluding it leaves 189 scored forecasts and an all-match score of 0.6138, with the paired comparison unchanged.

Janus has 1,533 scored forecasts across 186 matches. Its 1,368 abstentions are category-level entries across 152 fixtures. It also has 140 baseline-only entries, excluded from full-forecast scoring, and 21 pending entries; two baseline-only entries also appear in pending. Missing results or first-goal evidence were not guessed. Near-even forecasts remain included.

Same-forecast comparisons. Each comparison uses only rows with a saved benchmark. Sirius has 127 paired matches; the saved market probabilities have the margin removed. Individual quote timestamps are absent from the saved strip, so this article makes no exact T−30 or closing-price claim. Janus compares with the competition baseline saved at forecast time. Both team-scoring rows have 184 paired forecasts; first-to-score has 181. The chart shows these paired counts.

Scoring. Sirius averages the sum of the three squared probability errors per match (range 0–2). Janus binary categories average (probability − outcome)². First-to-score follows the tracker’s mean across three component errors. The Janus chart shows 100 × (1 − model error / baseline error). Its positive values mean lower error. These are descriptive comparisons; raw scores should not be ranked across engines or categories.

Outcome rules. Sirius uses recorded full-time scores where available. Janus preserves the tracker’s FT-only and score sanity checks. Cards count recorded yellow plus red cards, one each; cards and corners require both team-stat rows. First-to-score additionally requires consistent goal-event evidence. Extra-time/penalty cases that fail the tracker rules remain unscored.

Example selection. Betis–Madrid had the largest favourable Sirius-versus-market error difference. Schalke–Bayern had the largest unfavourable Janus binary-forecast difference against the league baseline. Both examples are selected after results and illustrate scoring; they are not independent tests of model skill.

Limits. One week is too short to establish a lasting advantage. Forecast categories within the same match are related. Scores measure probability quality, not financial returns. Case-study percentages are independently rounded to one decimal and can total 100.1%; calculations use the saved full precision. Future result corrections should be dated, not silently substituted into this edition.

JANUS PAIRED COMPARISONS
Values below use full saved calculation precision; the article rounds for readability.

Goals · 2.5: n=186; Janus Brier=0.2363526780; baseline Brier=0.2385135753; relative error reduction=+0.905985%.
Goals · 1.5: n=186; Janus Brier=0.1806556242; baseline Brier=0.1833146505; relative error reduction=+1.450526%.
Goals · 3.5: n=186; Janus Brier=0.2392098397; baseline Brier=0.2394879032; relative error reduction=+0.116108%.
Both teams score: n=186; Janus Brier=0.2514761266; baseline Brier=0.2499559140; relative error reduction=-0.608192%.
Home team scores: n=184; Janus Brier=0.1703460658; baseline Brier=0.1856489130; relative error reduction=+8.242896%.
Away team scores: n=184; Janus Brier=0.2096222137; baseline Brier=0.2101211957; relative error reduction=+0.237473%.
Cards · 4.5: n=113; Janus Brier=0.2167323373; baseline Brier=0.2280351770; relative error reduction=+4.956621%.
Corners · 9.5: n=121; Janus Brier=0.2504765460; baseline Brier=0.2486828512; relative error reduction=-0.721278%.
First team to score: n=181; Janus Brier=0.1937914289; baseline Brier=0.1947430828; relative error reduction=+0.488671%.

Public tracker and methodology references:
https://www.oddsivio.com/sirius/2026
https://www.oddsivio.com/janus/2026
https://www.oddsivio.com/predictions/methodology

Evidence read times: Sirius 13 September 2026 21:03:04 UTC; Janus 13 September 2026 21:05:01 UTC. Both were 14 September in Europe/Bucharest. This is a fixed edition; later tracker updates do not silently replace these numbers.