Registered experiment: 600 fresh Kalshi pairs across 60 dates →Compare TORK, the simple average, and calibration →

CURRENT EXPERIMENT · LIVE STATUS

What have we proved?

Checked Oct 17, 11:46 AM CDT

The audit record is growing. A prediction advantage and customer demand remain unproven. The September checkpoint below is fixed; it is not the latest experiment result.

PRIMARY MARKET TEST422 / 60031 / 60 event dates · 4 awaiting results

COLLECTING NEW EVIDENCE

Inspect the fixed test →

METHOD COMPARISON419 / 60031 / 60 event dates

TORK versus a simple average, fixed calibration, and Kalshi. Model contribution scores stay hidden until the planned checkpoint.

Inspect methods and model contribution →

PRIMARY TEST EXCLUSIONS79135.0% enrolled among enrolled + excluded records

These records already have a house lock. This rate does not count every scheduled event. Missing forecasts are counted separately below.

Why records cannot enter the primary test

missing kalshi lock584
no directional forecast181
stale or late market15
ambiguous outcome mapping7
checkpoint timing mismatch4

Exclusions remain visible. Historical forecasts and missing prices are not created after an event.

Forecast coverage · last 7 days

One count per imported competition that has started, excluding cancellations. A missing pick stays in the denominator. A sport with no forecast pipeline is marked schedule only. Events absent from our source database cannot be counted here.

Sport	Imported	Pregame picks	Four models	Audited	Evidence holds	Missing pregame record
Tennis	261	141	141	140	118	2
College football	59	54	54	42	5	0
NHLSchedule only	51	0	0	0	0	51
NBASchedule only	42	0	0	0	0	42
MMA / UFC	17	17	17	16	0	0
NFL	15	14	14	13	1	0
MLB	9	2	2	2	7	0

Window: 2026-10-10 to 2026-10-17 UTC. Each row uses its latest valid directional pregame decision. “Four models” requires one valid report from each provider at that same lock. “Evidence holds” includes locked holds and reasoned WITHHELD records created and published before start. They remain outside pick coverage. “Audited” is a subset of pregame picks, not a coverage score.

More records alone do not prove an edge. The registered test needs both 600 matched outcomes and 60 distinct dates, plus the complete stopping day and outstanding results. Better probability scores would still require a separate test of trading costs and customer demand.

Loading current experiment and coverage…

TORK INTELLIGENCE LEDGER

278 frozen outcomes.
No victory lap.

This fixed September checkpoint supports audit and comparison. It does not establish a prediction advantage. Use the current experiment status above for progress.

FROZEN

First 200+ sports validation checkpointSep 14, 2026, 6:07 PM PDT · Prospective records only · No retrospective filling

SEPTEMBER CHECKPOINT VERDICTNo alpha proven yet

TORK.BROS was correct on 193 of 278 frozen decisions. The 69.4% headline is real, but it is not evidence that the system beats models or markets.

FROZEN OUTCOMES278first checkpoint

LIVE SPORTS AUDITS1446all sports · 21 awaiting audit

ELIGIBLE PREGAME LOCKS146798.6% audited

MODEL-PAIRED277same events, five systems

ENSEMBLE-PAIRED126same events, six systems

Live totals use the same record as the homepage: eligible PICK and LEAN forecasts locked before the event, with completed audits counted after an official result. The 278 outcomes frozen on Sep 14, 2026, 6:07 PM PDT are included in the live total. Comparison scores below use that frozen subset.

LIVE COUNTS BY SPORT1446 completed audits · 1467 eligible pregame locks+

MLB283 completed audits

285 eligible pregame locks · 2 awaiting audit

MMA70 completed audits

71 eligible pregame locks · 1 awaiting audit

Boxing9 completed audits

9 eligible pregame locks · 0 awaiting audit

NFL70 completed audits

72 eligible pregame locks · 2 awaiting audit

College Football343 completed audits

358 eligible pregame locks · 15 awaiting audit

College Basketball0 completed audits

0 eligible pregame locks · 0 awaiting audit

Tennis671 completed audits

672 eligible pregame locks · 1 awaiting audit

NBA0 completed audits

0 eligible pregame locks · 0 awaiting audit

WNBA0 completed audits

0 eligible pregame locks · 0 awaiting audit

NHL0 completed audits

0 eligible pregame locks · 0 awaiting audit

OPERATIONAL MILESTONE

1446 of 1000 verified outcomes

100%

THE FOUR TESTS

What the data proves—and what it does not

Every comparison fails closed when a forecast, lock, probability, or official result is missing.

Forecasting alphaNOT PROVEN

Gemini versus TORK.BROS: +0.7 pts in accuracy.

277 events contain TORK.BROS and all four models on the same outcomes. The differences are too small to show a durable edge.

Market alphaNOT PROVEN

Compare probabilities on the same events.

137 Kalshi pairs and 144 Polymarket pairs qualify. Lower Brier is better. These fixed comparisons are development evidence, not a result from the registered test.

Ensemble alphaINSUFFICIENT DATA

Ensemble v2 is 66.7% vs 66.7% for TORK.BROS.

126 exact six-system outcomes qualify. The ensemble has not improved accuracy or calibration.

Routing alphaNOT TESTED

Only 8 events had model-side disagreement.

That is too little signal for a routing claim. A routing rule must be fixed first, then tested on an untouched holdout cohort.

WHAT THE CHECKPOINT SHOWS

The audit loop has produced a testable record.

TORK.BROS can timestamp forecasts, lock them before events, preserve the source record, resolve official outcomes, and score each system on the same event set. That verified decision history is the current asset.

278immutable cohort members

IDENTICAL MODEL COHORT

Who is winning on the same events?

277 included · 1 excluded · lower Brier and log loss are better

SourceRecordAccuracyBrierLog loss

CURRENT LEADERGemini195–8270.4%0.1870.538

FRONTIER MODELChatGPT194–8370.0%0.1910.554

HOUSETORK.BROS193–8469.7%0.1870.542

FRONTIER MODELGrok192–8569.3%0.1880.547

FRONTIER MODELClaude191–8669.0%0.1870.538

These are descriptive scores on the frozen cohort. A small difference in accuracy alone does not prove a model-selection advantage or independent reasoning.

MARKET CONTROL GROUP

Hit rate is not calibration

Each market uses its own exact paired cohort.

SAME-EVENT PAIR

TORK.BROS vs Kalshi

137 events

TORK.BROS65.0%Brier 0.223

Kalshi63.5%Brier 0.208

SAME-EVENT PAIR

TORK.BROS vs Polymarket

144 events

TORK.BROS59.7%Brier 0.236

Polymarket57.6%Brier 0.228

Accuracy counts winners. Brier measures the quality of probabilities. A higher hit rate alone does not establish market value; the fixed prospective test above is the primary test.

DISAGREEMENT SIGNAL

There is not enough routing evidence

277 events with all four model opinions

SAME SIDE, LOW SPREAD229models mostly agreed

SAME SIDE, 10+ PT SPREAD40confidence differed

DIFFERENT WINNERS850.0% TORK accuracy

A routing rule must be fixed before a new test. Selecting a rule after inspecting these outcomes would contaminate the test. Agreement between providers does not prove independent evidence.

FROZEN CHECKPOINT BY SPORT

Why the sport mix matters

This dated subset stays fixed as live audit totals grow.

MLB136 graded

56.6%

77 correct · 59 wrong

College Football82 graded

87.8%

72 correct · 10 wrong

MMA25 graded

72.0%

18 correct · 7 wrong

Tennis15 graded

66.7%

10 correct · 5 wrong

NFL11 graded

72.7%

8 correct · 3 wrong

Boxing9 graded

88.9%

8 correct · 1 wrong

Different sports have different favorite strengths. Compare each method with the same-event market before drawing a conclusion from a combined accuracy rate.

IMMUTABLE LEDGER

Inspect the frozen decisions

Open all results →

CORRECTSep 14

Nao Hibino vs Magdalena FrechTennis · Winner: Magdalena Frech

TORK.BROS PICKMagdalena Frech58.8%

TORK.BROSMagdalena Frech58.8% · correct

KalshiMagdalena Frech79.0% · correct

ClaudeMagdalena Frech61.0% · correct

GeminiMagdalena Frech58.0% · correct

GrokMagdalena Frech58.0% · correct

ChatGPTMagdalena Frech58.0% · correct

Forecast coverage · last 7 days

278 frozen outcomes.No victory lap.

The audit loop has produced a testable record.

Live calibration

MLB data conditions

Historical reconstruction

Forecast coverage · last 7 days

278 frozen outcomes.
No victory lap.