Soccer prediction system deep audit

Slate 20261106 · immutable forensic review of 2 match prediction artifact(s)

Executive conclusion

In 2 match(es) the headline score is not the distribution's modal scoreline (a breach of the published-mode contract) — see the MOST_LIKELY_IS_MODE findings.

At a glance

2matches audited
0critical
0errors
3warnings
0grid non-reproducible
2headline != modal
0winner disagreements
0lambda inconsistencies
0with real result
-winner hit-rate
-decisive hit-rate
0%predicted draws
-actual draws
0confident misses
-mean Brier
 

Matches

Everton vs Coventry

20261106_COC@EVE · Premier League
WARNINGS risk 20 pred 1-0 (Everton)λ 1.77 / 1.01conf 0.69 warning 2 good 39
Problems (2)
SevCategoryRuleMessageValueExpected
warningmodel_behaviorMOST_LIKELY_IS_MODEHeadline score is not the modal scoreline (modal=1-1 (p=0.1118), shown=1-0 (p=0.1088)); the published exact score must be the grid mode, and the winning scenario belongs in the labelled most_likely_winning_score_home/away fields. Both lean Everton.1-01-1
warningprovenanceFBREF_PRIORS_ACTIVEFBref dataset files not found; team priors inactive.
Lambda evolution (audit trail)
StageHome λAway λ
Pre-competitive-floor1.6570.838
Pre-calibration1.6810.826
Final1.7681.013
competition_calibration_scale: 1.115competition_observed_total_goals: 2.781quality_gap_applied: Truemismatch_score_pressure_applied: Falsecompetitive_underdog_floor_applied: Falsestage_discount: 1.0
Scoreline distribution
Top scorelinesP
1-111.2%
1-010.9%
2-19.8%
2-09.7%
0-06.3%
Outcome — home 55.0% · draw 23.5% · away 21.4%
Markets — O2.5 52.6% · BTTS 52.9%
Passed checks (39)
SevCategoryRuleMessageValueExpected
goodintegritySCHEMA_KEYS_PRESENTAll required sections present.
goodstructureLINEUP_ELEVENEverton: 11 starters.
goodstructureFORMATION_VALIDEverton: formation 4-2-3-1 valid.
goodstructureLINEUP_ELEVENCoventry: 11 starters.
goodstructureFORMATION_VALIDCoventry: formation 4-2-3-1 valid.
goodprobabilityOUTCOME_SUMS_ONEoutcome sums to 1.0000.
goodprobabilityMARKETS_COMPLEMENTARYOver/Under and BTTS pairs each sum to 1.
goodprobabilitySCORELINE_PROBS_NONNEGAll scoreline probabilities within [0,1].
goodprobabilityGRID_SUMS_ONEscoreline grid sums to 0.9972.
goodmodel_behaviorGRID_REPRODUCIBLEGrid reproducible from lambdas (max cell diff 0.0008).
goodprobabilityMLS_PROBABILITY_MATCHESHeadline probability matches the grid.
goodmodel_behaviorOUTCOME_MATCHES_GRIDOutcome split matches the grid (max diff 0.0020).
goodmodel_behaviorMARKETS_MATCH_GRIDMarkets match the grid (max diff 0.0029).
goodintegrityWINNER_IS_VALID_TEAMpredicted_winner 'Everton' is a participating side.
goodmodel_behaviorWINNER_MATCHES_SCOREWinner agrees with headline score (Everton).
goodmodel_behaviorWINNER_PROBABILITY_DIRECTIONpredicted_winner 'Everton' follows the published modal scoreline 1-0.
goodmodel_behaviorLAMBDA_BOUNDSLambdas in range (home 1.768, away 1.013).
goodprovenanceLAMBDA_FACTORS_PRESENTlambda_factors audit trail present.
goodmodel_behaviorLAMBDA_CONSISTENCYLambda identical across prediction / antagonist / factors.
goodmodel_behaviorCALIBRATION_SCALE_BOUNDSCalibration scale 1.115 in range.
goodprovenanceCALIBRATION_RECORDEDCalibration recorded with pre/post lambdas.
goodprovenanceXGBOOST_REASONING_PRESENTMatch-level xgboost_reasoning present (XGBoost).
goodmodel_behaviorLAMBDA_REASONING_CONSISTENCYLambdas in xgboost_reasoning match final Poisson lambdas (1.768 / 1.013).
goodprovenanceXGBOOST_PLAYER_REASONINGAll 22 starters carry full XGBoost / FBref reasoning.
goodintegrityTEAM_GOALS_SUMEverton: player goals reconcile to team total (1).
goodintegrityTEAM_GOALS_SUMCoventry: player goals reconcile to team total (0).
goodstructurePOSSESSION_SUMS_100Possession sums to 100.0%.
goodintegrityPLAYER_STAT_REALISMAll 22 projected player lines pass realism checks.
goodstructureEXACTLY_ONE_GOALKEEPEREverton: exactly one goalkeeper (Jordan Pickford).
goodintegritySHOT_MATH_INTEGRITYAll 11 projected lines have consistent shot components.
goodintegrityTEAM_SHOTS_CAPEverton: 15 projected shots inside the realistic envelope.
goodstructureEXACTLY_ONE_GOALKEEPERCoventry: exactly one goalkeeper (Carl Rushworth).
goodintegritySHOT_MATH_INTEGRITYAll 11 projected lines have consistent shot components.
goodintegrityTEAM_SHOTS_CAPCoventry: 10 projected shots inside the realistic envelope.
goodprovenanceSTANDING_AVAILABLELeague-standing context available for both sides.
goodprovenanceRANKING_PROVENANCE22/22 starters carry a ranking source.
goodprovenanceDATA_DEPTH22/22 starters have adequate recent history.
goodprovenancePREGAME_DIRECTION_AGREESPregame favorite agrees with the model winner (Everton).
goodprovenancePREGAME_TOTAL_CONSISTENTPregame total 2.78 ~ Poisson total 2.78.

Sevilla vs Alavés

20261106_ALA@SEV · LaLiga
WARNINGS risk 10 pred 2-1 (Sevilla)λ 1.60 / 1.38conf 0.66 warning 1 good 40
Problems (1)
SevCategoryRuleMessageValueExpected
warningmodel_behaviorMOST_LIKELY_IS_MODEHeadline score is not the modal scoreline (modal=1-1 (p=0.1128), shown=2-1 (p=0.0898)); the published exact score must be the grid mode, and the winning scenario belongs in the labelled most_likely_winning_score_home/away fields. Both lean Sevilla.2-11-1
Lambda evolution (audit trail)
StageHome λAway λ
Pre-competitive-floor1.4931.447
Pre-calibration1.5621.297
Final1.6041.379
competition_calibration_scale: 1.015competition_observed_total_goals: 2.984quality_gap_applied: Truemismatch_score_pressure_applied: Falsecompetitive_underdog_floor_applied: Falsestage_discount: 1.0
Scoreline distribution
Top scorelinesP
1-111.3%
2-19.0%
1-08.0%
1-27.7%
0-16.9%
Outcome — home 42.8% · draw 24.4% · away 32.9%
Markets — O2.5 57.3% · BTTS 59.9%
Passed checks (40)
SevCategoryRuleMessageValueExpected
goodintegritySCHEMA_KEYS_PRESENTAll required sections present.
goodstructureLINEUP_ELEVENSevilla: 11 starters.
goodstructureFORMATION_VALIDSevilla: formation 4-2-3-1 valid.
goodstructureLINEUP_ELEVENAlavés: 11 starters.
goodstructureFORMATION_VALIDAlavés: formation 5-3-2 valid.
goodprobabilityOUTCOME_SUMS_ONEoutcome sums to 1.0000.
goodprobabilityMARKETS_COMPLEMENTARYOver/Under and BTTS pairs each sum to 1.
goodprobabilitySCORELINE_PROBS_NONNEGAll scoreline probabilities within [0,1].
goodprobabilityGRID_SUMS_ONEscoreline grid sums to 0.9965.
goodmodel_behaviorGRID_REPRODUCIBLEGrid reproducible from lambdas (max cell diff 0.0009).
goodprobabilityMLS_PROBABILITY_MATCHESHeadline probability matches the grid.
goodmodel_behaviorOUTCOME_MATCHES_GRIDOutcome split matches the grid (max diff 0.0017).
goodmodel_behaviorMARKETS_MATCH_GRIDMarkets match the grid (max diff 0.0035).
goodintegrityWINNER_IS_VALID_TEAMpredicted_winner 'Sevilla' is a participating side.
goodmodel_behaviorWINNER_MATCHES_SCOREWinner agrees with headline score (Sevilla).
goodmodel_behaviorWINNER_PROBABILITY_DIRECTIONpredicted_winner 'Sevilla' follows the published modal scoreline 2-1.
goodmodel_behaviorLAMBDA_BOUNDSLambdas in range (home 1.604, away 1.379).
goodprovenanceLAMBDA_FACTORS_PRESENTlambda_factors audit trail present.
goodmodel_behaviorLAMBDA_CONSISTENCYLambda identical across prediction / antagonist / factors.
goodmodel_behaviorCALIBRATION_SCALE_BOUNDSCalibration scale 1.015 in range.
goodprovenanceCALIBRATION_RECORDEDCalibration recorded with pre/post lambdas.
goodprovenanceXGBOOST_REASONING_PRESENTMatch-level xgboost_reasoning present (XGBoost).
goodmodel_behaviorLAMBDA_REASONING_CONSISTENCYLambdas in xgboost_reasoning match final Poisson lambdas (1.604 / 1.379).
goodprovenanceFBREF_PRIORS_ACTIVEFBref team priors applied (blend weight 0.228, home E[lambda]=2.474, away E[lambda]=2.469).
goodprovenanceXGBOOST_PLAYER_REASONINGAll 22 starters carry full XGBoost / FBref reasoning.
goodintegrityTEAM_GOALS_SUMSevilla: player goals reconcile to team total (2).
goodintegrityTEAM_GOALS_SUMAlavés: player goals reconcile to team total (1).
goodstructurePOSSESSION_SUMS_100Possession sums to 100.0%.
goodintegrityPLAYER_STAT_REALISMAll 22 projected player lines pass realism checks.
goodstructureEXACTLY_ONE_GOALKEEPERSevilla: exactly one goalkeeper (Odysseas Vlachodimos).
goodintegritySHOT_MATH_INTEGRITYAll 11 projected lines have consistent shot components.
goodintegrityTEAM_SHOTS_CAPSevilla: 13 projected shots inside the realistic envelope.
goodstructureEXACTLY_ONE_GOALKEEPERAlavés: exactly one goalkeeper (Antonio Sivera).
goodintegritySHOT_MATH_INTEGRITYAll 11 projected lines have consistent shot components.
goodintegrityTEAM_SHOTS_CAPAlavés: 10 projected shots inside the realistic envelope.
goodprovenanceSTANDING_AVAILABLELeague-standing context available for both sides.
goodprovenanceRANKING_PROVENANCE22/22 starters carry a ranking source.
goodprovenanceDATA_DEPTH22/22 starters have adequate recent history.
goodprovenancePREGAME_DIRECTION_AGREESPregame favorite agrees with the model winner (Sevilla).
goodprovenancePREGAME_TOTAL_CONSISTENTPregame total 2.98 ~ Poisson total 2.98.

Rule catalogue

RuleDescriptionTriggered
SCHEMA_KEYS_PRESENTAll required top-level sections exist in the artifact.ok
LINEUP_ELEVENEach side fields exactly 11 starters.ok
FORMATION_VALIDFormation string parses and its outfield lines sum to 10.ok
POSSESSION_SUMS_100Estimated possession of both teams sums to ~100%.ok
TEAM_GOALS_SUMPlayers' projected integer goals sum to the team's headline score.ok
TEAM_TOTALS_SCORE_MATCHteam_totals.projected_goals equals the headline score component.ok
TEAM_EG_LAMBDA_MATCHteam_totals.expected_goals matches the team's Poisson lambda.ok
PLAYER_STAT_REALISMProjected player lines are internally consistent (xG<=shots, accurate<=total passes, minutes/rating in range).ok
EXACTLY_ONE_GOALKEEPEREach starting XI fields exactly one goalkeeper.ok
SHOT_MATH_INTEGRITYShot components are physically consistent (on target <= total, on+off <= total; blocked shots cannot be negative).ok
PLAYER_MINUTES_LEAGUE_CAPNo player projects more minutes than the competition clock allows (90 + stoppage, or extra time in cups).ok
TEAM_SHOTS_CAPTeam projected shots stay inside the realistic historical envelope (~35).ok
OUTCOME_SUMS_ONEhome_win + draw + away_win sums to 1.ok
MARKETS_COMPLEMENTARYOver/Under and BTTS pairs each sum to 1.ok
GRID_SUMS_ONEThe scoreline probability grid sums to ~1.ok
SCORELINE_PROBS_NONNEGEvery scoreline probability is a valid [0,1] value.ok
GRID_REPRODUCIBLEThe reported scoreline grid is reproducible from the lambdas.ok
MOST_LIKELY_IS_MODEThe headline most-likely score is the distribution's modal scoreline.2 problems
MLS_PROBABILITY_MATCHESmost_likely_score_probability equals the grid value of that score.ok
OUTCOME_MATCHES_GRIDThe reported win/draw/loss split is derivable from the grid.ok
MARKETS_MATCH_GRIDThe reported Over/Under and BTTS are derivable from the grid.ok
WINNER_MATCHES_SCOREpredicted_winner agrees with the headline score direction.ok
WINNER_PROBABILITY_DIRECTIONpredicted_winner derives from the published modal scoreline; diverging from the 1X2 favorite is expected.ok
WINNER_IS_VALID_TEAMpredicted_winner names a participating team (or Draw).ok
MODAL_OUTCOME_IS_DRAWDraw is the single largest outcome bucket while a team is named.ok
LAMBDA_BOUNDSBoth lambdas fall inside the engine's [0.55, 5.00] envelope.ok
LAMBDA_CONSISTENCYThe lambda is identical across prediction / antagonist / lambda_factors.ok
LAMBDA_FACTORS_PRESENTThe lambda_factors decomposition (audit trail) is present.ok
CALIBRATION_SCALE_BOUNDScompetition_calibration_scale stays within [0.85, 1.35].ok
CALIBRATION_RECORDEDWhen calibration ran, the pre-calibration lambdas are recorded.ok
STANDING_AVAILABLELeague-standing context is available for a domestic fixture.ok
RANKING_PROVENANCEPlayer ranking/quality signals carry a recognised source.ok
DATA_DEPTHStarters have enough recent games to model rather than synthesise.ok
FBREF_PRIORS_ACTIVEFBref season-prior files exist so that branch is not silently inert.1 problems
XGBOOST_REASONING_PRESENTMatch-level xgboost_reasoning payload is present with full tracing pipeline.ok
XGBOOST_PLAYER_REASONINGStarters have xgboost_reasoning metadata tracking feature contributions and FBref priors.ok
LAMBDA_REASONING_CONSISTENCYLambdas reported in xgboost_reasoning match the final Poisson lambdas.ok
PREGAME_DIRECTION_AGREESThe pregame fair-line favorite agrees with the model winner.ok
PREGAME_TOTAL_CONSISTENTThe pregame expected total is close to the sum of the lambdas.ok
RESULT_WINNER_HITThe predicted winner matched the real result.ok
RESULT_EXACT_SCORE_HITThe headline score matched the real scoreline exactly.ok
RESULT_OU_HITThe Over/Under 2.5 lean matched reality.ok
RESULT_BTTS_HITThe BTTS lean matched reality.ok
RESULT_CONFIDENT_MISSA high-confidence prediction did not reach the real outcome.ok
OUTCOME_BRIERBrier score of the win/draw/loss probabilities vs the real result.ok
Generated 2026-11-03 14:38:27 UTC · Soccer prediction reporting