{"asset":"goalpulse-model-search-log","publisher":"GoalPulse","canonicalDomain":"https://goalpulse.io","url":"https://goalpulse.io/data/football-model-search.json","humanPage":"https://goalpulse.io/football/experiments","generatedAt":"2026-11-19T14:41:54.745Z","summary":{"rounds":60,"listedRounds":29,"priorRounds":31,"proposers":4,"unique":20,"duplicates":5,"rejected":4,"better":4,"equal":10,"worse":6,"beatValidationLostTest":1,"costUsd":59.31,"oneLine":"60 model proposals across 4 LLM families; 4 beat production on walk-forward validation and 1 of those also lost on the locked test, so production was unchanged."},"shipped":{"round":26,"validationRps":0.2039,"validationLogLoss":0.9976,"testRps":0.2051,"testLogLoss":1.0047,"baselines":{"resultsOnlyTestRps":0.2087,"marketPriorTestRps":0.2071}},"priorSearch":{"rounds":31,"proposer":"claude-fable-5-1","costUsd":24.65,"adoptedRound":14,"note":"The 2026-11-07 search produced the shipped plugin; rounds 15-30 found no significant gain."},"costByProposer":{"claude-fable-5-1":24.65,"gpt-5.1":0.67,"claude-fable-5-2":8.48,"gpt-6.1-sol":0.86},"shadowExperiment":{"name":"injury-intel","neverTouchesPublishedProbabilities":true,"adoptRule":"paired90ciExcludesZeroOnTheBetterSide","minScored":150,"publishesInterimResult":false,"note":"Scored alongside the published probabilities on the same matches; below the pre-registered threshold no interim difference is published."},"rounds":[{"round":31,"proposer":"gpt-5.1","hypothesis":"Rescale score-only goal observations so each family’s long-run training goals match its own mean of real goals.","validationRps":0.2039,"validationLogLoss":0.9976,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":32,"proposer":"gpt-5.1","hypothesis":"Replace the family-wide goal-level correction with a tier-aware level (tier-1 vs tier-2), blended by recent tier mix.","validationRps":0.2039,"validationLogLoss":0.9976,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":33,"proposer":"gpt-5.1","hypothesis":"Give the most recent match a small extra weight per refit so ratings react faster to genuine strength changes.","validationRps":0.2039,"validationLogLoss":0.9976,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":34,"proposer":"gpt-5.1","hypothesis":"Add a small symmetric rest-days term to the log-strength ratio.","validationRps":0.204,"validationLogLoss":0.998,"testRps":null,"testLogLoss":null,"verdict":"worse","rejectedReason":null,"overfitNote":null},{"round":35,"proposer":"gpt-5.1","hypothesis":"Reweight short-term form/tempo residuals to give market-observed matches slightly more effective samples.","validationRps":0.204,"validationLogLoss":0.998,"testRps":null,"testLogLoss":null,"verdict":"worse","rejectedReason":null,"overfitNote":null},{"round":36,"proposer":"gpt-5.1","hypothesis":"Taper the favourite stretch and rho shift in the first ten matchweeks.","validationRps":null,"validationLogLoss":null,"testRps":null,"testLogLoss":null,"verdict":"contract-rejection","rejectedReason":"plugin longer than 20,000 characters","overfitNote":null},{"round":37,"proposer":"gpt-5.1","hypothesis":"Ship the round-26 model unchanged.","validationRps":null,"validationLogLoss":null,"testRps":null,"testLogLoss":null,"verdict":"contract-rejection","rejectedReason":"plugin longer than 20,000 characters","overfitNote":null},{"round":31,"proposer":"claude-fable-5-2","hypothesis":"Opponent-adjust the fast form and tempo layers by backfitting, with the shrinkage re-derived each pass.","validationRps":0.2039,"validationLogLoss":0.9976,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":32,"proposer":"claude-fable-5-2","hypothesis":"Adjust the form layer for rest asymmetry.","validationRps":null,"validationLogLoss":null,"testRps":null,"testLogLoss":null,"verdict":"contract-rejection","rejectedReason":"no <plugin> block in the answer","overfitNote":null},{"round":33,"proposer":"claude-fable-5-2","hypothesis":"Scale the tier-gap correction by the priced share of the current tier-2 teams’ training evidence.","validationRps":0.2039,"validationLogLoss":0.9976,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":34,"proposer":"claude-fable-5-2","hypothesis":"Count the season break as one extra in-season half-life for priced rows.","validationRps":0.204,"validationLogLoss":0.9979,"testRps":null,"testLogLoss":null,"verdict":"worse","rejectedReason":null,"overfitNote":null},{"round":35,"proposer":"claude-fable-5-2","hypothesis":"Split the market residual by league tier.","validationRps":null,"validationLogLoss":null,"testRps":null,"testLogLoss":null,"verdict":"contract-rejection","rejectedReason":"no <plugin> block in the answer","overfitNote":null},{"round":36,"proposer":"claude-fable-5-2","hypothesis":"Scale the tier-gap correction by the mean priced share of the family.","validationRps":0.2039,"validationLogLoss":0.9976,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":37,"proposer":"claude-fable-5-2","hypothesis":"Make the rho shift depend on match closeness.","validationRps":0.2039,"validationLogLoss":0.9978,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":38,"proposer":"gpt-6.1-sol","hypothesis":"Preserve the predicted goal total when applying strength, tier and form corrections.","validationRps":0.2039,"validationLogLoss":0.9978,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":39,"proposer":"gpt-6.1-sol","hypothesis":"Use only priced observations in the fast form and tempo layers, so short-window realised scores cannot add noise to unpriced fixtures.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":0.2053,"testLogLoss":1.0048,"verdict":"beat-validation","rejectedReason":null,"overfitNote":"validation −0.0001 RPS, locked test +0.0002 RPS versus production (0.2051 / 1.0047)"},{"round":40,"proposer":"gpt-6.1-sol","hypothesis":"Consolidate round 39 unchanged, so no further validation-tuned parameter is added.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"beat-validation","rejectedReason":null,"overfitNote":null},{"round":41,"proposer":"gpt-6.1-sol","hypothesis":"Remove only the fast tempo adjustment and keep everything else.","validationRps":0.2038,"validationLogLoss":0.9976,"testRps":null,"testLogLoss":null,"verdict":"beat-validation","rejectedReason":null,"overfitNote":null},{"round":42,"proposer":"gpt-6.1-sol","hypothesis":"Restore the market-only fast tempo adjustment on top of round 39.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"beat-validation","rejectedReason":null,"overfitNote":null},{"round":43,"proposer":"gpt-6.1-sol","hypothesis":"Remove only the fast strength-form adjustment, keeping tempo.","validationRps":0.2047,"validationLogLoss":1.0002,"testRps":null,"testLogLoss":null,"verdict":"worse","rejectedReason":null,"overfitNote":null},{"round":44,"proposer":"gpt-6.1-sol","hypothesis":"Restore market-only fast strength form while keeping tempo.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"same-as-round-39","rejectedReason":null,"overfitNote":null},{"round":45,"proposer":"gpt-6.1-sol","hypothesis":"Train priced rows on market-implied goals alone instead of the 75/25 mixture.","validationRps":0.2042,"validationLogLoss":0.9983,"testRps":null,"testLogLoss":null,"verdict":"worse","rejectedReason":null,"overfitNote":null},{"round":46,"proposer":"gpt-6.1-sol","hypothesis":"Restore the 75% market / 25% compressed-score training mixture.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"same-as-round-39","rejectedReason":null,"overfitNote":null},{"round":47,"proposer":"gpt-6.1-sol","hypothesis":"Remove blow-out compression and its compensating goal-level correction.","validationRps":0.2039,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":48,"proposer":"gpt-6.1-sol","hypothesis":"Restore blow-out compression and its goal-level correction.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"same-as-round-39","rejectedReason":null,"overfitNote":null},{"round":49,"proposer":"gpt-6.1-sol","hypothesis":"Remove only the manual rho offset.","validationRps":0.2039,"validationLogLoss":0.9978,"testRps":null,"testLogLoss":null,"verdict":"tied","rejectedReason":null,"overfitNote":null},{"round":50,"proposer":"gpt-6.1-sol","hypothesis":"Restore the source-aware rho offset.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"same-as-round-39","rejectedReason":null,"overfitNote":null},{"round":51,"proposer":"gpt-6.1-sol","hypothesis":"Remove only the tier-gap correction.","validationRps":0.2042,"validationLogLoss":0.9986,"testRps":null,"testLogLoss":null,"verdict":"worse","rejectedReason":null,"overfitNote":null},{"round":52,"proposer":"gpt-6.1-sol","hypothesis":"Restore the source-scaled tier correction and otherwise ship round 39.","validationRps":0.2038,"validationLogLoss":0.9975,"testRps":null,"testLogLoss":null,"verdict":"same-as-round-39","rejectedReason":null,"overfitNote":null}],"doNotClaim":["A validation-season win is not a shipped improvement; only the locked test decides, and this log shows one case where the sign flipped.","These are model-architecture experiments on historical matches; they say nothing about any individual fixture.","No round in this log promises a win rate, an edge over bookmakers, or a return."],"caveat":"Proposals were generated by LLMs but every score is produced by a deterministic walk-forward harness on fixed data; no number here was written by a model."}