OUR FORECASTS, REVIEWED — METHODS 14–20 September 2026 Fixed edition. Evidence was retrieved at the cutoffs below. WINDOW AND CUTOFFS Kickoffs from 14 September 2026 00:00 inclusive to 21 September 2026 00:00 exclusive, Europe/Bucharest (EEST, UTC+3). UTC bounds: 13 September 21:00 inclusive to 20 September 21:00 exclusive. Assignment uses saved ledger kickoff_at. A Sunday evening fixture in the Americas can therefore belong to Monday locally and be outside this edition. Sirius database read timestamp: 2026-09-21T08:27:40.888531+00:00 (11:27:40.888531 EEST). Janus repeatable-read transaction timestamp: 2026-09-21T08:27:46.988946+00:00 (11:27:46.988946 EEST). Sirius uses sequential read-only SELECTs, not a shared repeatable-read snapshot; Janus uses one read-only repeatable-read transaction. These are actual retrieval times, not a claim of a 09:00 snapshot. The immutable local exports retain results/statistics/events observed then. Later corrections need a dated amendment. FORECAST INTEGRITY Only frozen records qualify. All scored Sirius rows have generation, creation and evidence-cutoff timestamps before kickoff; all 220 have matching saved/current team orientation and kickoff. Janus's frozen records likewise have no missing or post-kickoff generation, creation or evidence-cutoff timestamps. frozen_at equals the logical kickoff boundary by service design: it is not an independent wall-clock record of when the freezing job ran. No forecasts were regenerated or production rows changed for this article. Sirius scored versions: ledger-v2.8.0 (15), ledger-v2.8.2 (13), ledger-v2.8.4 (120), ledger-v2.8.5 (72). Janus frozen records: 2.10.0. The pilot used earlier versions. This is a description of the forecasts actually stored, not an experiment isolating the effect of a model change. SIRIUS COVERAGE AND SCORING The current database contains 357 fixtures in the window and 357 frozen, serve-eligible Sirius ledgers. Of these, 136 have no forecast: 127 insufficient-track-record and 9 team-data-void abstentions. Another forecast, Levante–Athletic Club (115529), remains postponed. The other two postponed fixtures already sit inside the no-forecast count. Thus 220 forecasts are scored, with no orientation holds and none unfrozen. Eight scored forecasts lack a valid saved market strip; 212 are compared with the market. Sirius uses the mean, across matches, of the SUM of the three squared probability errors: sum((p_side - I(actual_side))²). Range 0–2. Every probability counts; near-even forecasts remain included. Outcomes use recorded full-time scores when available, falling back to recorded scores, under the current Sirius tracker rules. All-match Brier: 0.5830983585839447. Paired Sirius: 0.5855623555561601. Paired saved market: 0.5876701581889986. Relative error reduction: 0.3586710339919308%, rounded to 0.4%. Both ranked the result first in 111/212 paired matches (Sirius 116/220 across all scored matches). Saved market probabilities have the margin removed. The strips contain probabilities, book counts, displays, spreads and opening probabilities but no individual quote timestamps; this article makes no exact T−30 or closing-price claim. Both sides of the comparison use the same 212 matches. The 61 top-five-domestic-league matches are only part of the 220; no top-five subset is substituted for the full record. JANUS COVERAGE AND SCORING Seven physical ledgers expand to nine forecast categories. There are 2,499 physical rows; 2,417 are frozen, spanning 357 fixtures. Their 3,131 category entries consist of 1,820 scored full forecasts across 214 matches, 1,250 abstentions, 41 baseline-only entries with outcomes (excluded), and 20 pending entries. Baseline-only coverage totals 42 because one baseline-only cards entry is also pending. There are zero void entries. These categories are overlapping forecasts, not 1,820 independent matches; first-to-score contributes three component errors, bringing component-claim count to 2,130. Six categories have 214 paired forecasts each. Cards has 187, corners 194, and first-to-score 155. All scored forecasts have their complete saved baseline for comparison. Near-even forecasts remain included. Binary lanes use mean((p - outcome)²), range 0–1. First-to-score follows the tracker and averages the three component squared errors within each match, then across matches (maximum 2/3). It is not directly comparable with Sirius's summed score. Chart reduction = 100 × (1 - Janus Brier / saved-baseline Brier), calculated at full precision and rounded to one decimal. Positive means lower error; it is not a percentage-point accuracy gain. No pooled Janus score or cross-engine ranking is used. The baseline is the competition prior saved at forecast time in payload.anchor. Janus preserves the public tracker's FT-only, score-sanity, two-team-stat-row and first-goal-event guards. Cards count recorded yellow plus red cards, one each. Missing red is treated as zero by the tracker; missing yellow or corners makes the relevant outcome unavailable. First-to-score needs score-consistent current goal events, valid teams and no event minute beyond 90. Missing evidence is not guessed. The 20 pending category entries are 9 for Peterborough–Barnsley (121626, penalty-shootout status, excluded by the FT-only rule), 9 for postponed Levante–Athletic Club (115529), and 2 cards entries for missing statistics: Bayer Leverkusen–Celje (121799, baseline-only) and Houston–Cincinnati (44230, full). Pending means ungraded under these rules; it does not imply every entry will later settle. FIRST-TO-SCORE COVERAGE LIMITATION All 82 unfrozen physical rows belong to Sunday 20 September: 54 full forecasts and 28 abstentions. They remain outside both the score and frozen coverage counts. All had pre-kickoff generated timestamps, but generation alone does not satisfy the frozen-record rule. We did not freeze, backfill or score them. The 6.2% reduction describes only 155 qualifying matches and is explicitly labelled incomplete in the article/chart. The cause of the missing freeze is not established by this editorial extraction. Investigate separately before treating this as complete weekly coverage. EXAMPLES AND SENSITIVITY The pilot's rules are retained: choose the lowest single-match Sirius-minus-market Brier difference and the highest Janus binary-forecast-minus-baseline difference. These are retrospective explanatory selections, not independently chosen tests of skill. Lillestrøm–Torreense (121815): 17 September, 1–2. Sirius H/D/A: 0.27241308882036164 / 0.25652340425531917 / 0.4710635069243192. Saved market away probability: 0.16539634266918946. Sirius Brier 0.41978696159858786 versus market 1.1318745365294551. Removing it leaves 211 matches: Sirius 0.5863480209303666, market 0.5850909905191386. This sensitivity check illustrates fragility; the published headline retains all 212. Grêmio–Palmeiras (43742): 20 September, 0–0. Janus away-team scoring 0.8847; frozen league baseline 0.705. Binary error 0.78269409 versus 0.497025. Both independently rounded display percentages and full saved values are retained. We make no tactical explanation for the miss. VALIDATION AND SOURCES The dated public API snapshots match all Sirius scored/paired IDs and hit flags and all Janus scored category IDs and binary displayed probabilities. Scores were independently recalculated from retained per-forecast values. Calculations follow the current public tracker grading rules. Primary records: https://oddsivio-production.up.railway.app/api/sirius/2026-tracker https://oddsivio-production.up.railway.app/api/janus/2026-tracker https://www.oddsivio.com/sirius/2026 https://www.oddsivio.com/janus/2026 https://www.oddsivio.com/predictions/methodology https://www.oddsivio.com/matches/121815 https://www.oddsivio.com/matches/43742 Independent result corroboration (accessed 21 September 2026): Liga Portugal, Lillestrøm 1–2 Torreense, 17 September: https://www.ligaportugal.pt/match/20262027/219039/1/12627746 Palmeiras official dated match album (Grêmio 0x0 Palmeiras): https://www.palmeiras.com.br/central-de-midia/fotos/album/?n=741d0507-1906-5037-a73a-b47f41b7c545&t=gremio-0x0-palmeiras Additional independent reporting confirms the 0–0 score; no tactical claims are imported into the article: https://www1.folha.uol.com.br/esporte/2026/09/palmeiras-perde-gols-e-fica-no-zero-com-o-gremio-em-porto-alegre.shtml LIMITS One week cannot establish durable skill. Related category forecasts are not independent trials. Model versions, competition mix and coverage vary. The 31 August–6 September pilot is the comparison edition, not the intervening 7–13 September week. Probability-error reduction is not a financial-return claim. This fixed edition preserves the retrieval cutoffs; current live trackers can change as results or freeze status change. JANUS PAIRED COMPARISONS (full calculation precision) Goals · over/under 2.5: n=214; Janus Brier=0.2388163525; baseline Brier=0.2402199766; reduction=+0.584308% Goals · over/under 1.5: n=214; Janus Brier=0.1655770242; baseline Brier=0.1679418224; reduction=+1.408106% Goals · over/under 3.5: n=214; Janus Brier=0.2317324832; baseline Brier=0.2355147196; reduction=+1.605945% Both teams to score: n=214; Janus Brier=0.2428930991; baseline Brier=0.2424274533; reduction=-0.192076% Home team to score: n=214; Janus Brier=0.1508348927; baseline Brier=0.1583356308; reduction=+4.737240% Away team to score: n=214; Janus Brier=0.1954410200; baseline Brier=0.2023634346; reduction=+3.420783% Cards · over/under 4.5: n=187; Janus Brier=0.2396165117; baseline Brier=0.2427422460; reduction=+1.287676% Corners · over/under 9.5: n=194; Janus Brier=0.2466949595; baseline Brier=0.2477590206; reduction=+0.429474% First team to score*: n=155; Janus Brier=0.1668610728; baseline Brier=0.1779607103; reduction=+6.237128%