← Results and first checkpointREGISTERED EXPERIMENT 422 / 600Graded event pairs 31 / 60Distinct event dates · UTC 4Enrolled pairs awaiting a valid result 791Started events excluded
Can TORK improve on the market?
We will compare new forecasts with Kalshi on the same events, at the same cutoff. Earlier results remain development evidence.
COLLECTING NEW EVIDENCEThe rules are fixed
- Primary test
- TORK Brier score minus Kalshi Brier score on each event. A negative difference favors TORK.
- Entry rule
- One pair per competition, enrolled before the start. Both source records must exist at the forecast cutoff. Market quotes must be no more than 15 minutes old.
- Version control
- House, model, and prompt versions are recorded at registration. Changed versions require a new experiment.
- Analysis checkpoint
- At least 600 graded pairs across 60 event dates. Include all enrolled events through that UTC day and wait for outstanding outcomes.
- Success rule
- Average Brier improvement of at least 0.01, with the upper end of a 95% confidence interval below zero. Resample whole dates to account for shared conditions.
- During collection
- Show counts, coverage, and exclusions. Do not declare success early. The 1,000-result counter is a separate operational milestone.
- Separate evidence
- World events, Polymarket, sport breakdowns, model agreement, and Ensemble v2 are secondary analyses. They do not replace the primary test.
Coverage and exclusions
Missing or late data stays excluded. The system does not reconstruct forecasts after an event.
Graded pairs by sport
MLB 80
MMA 16
NFL 29
TENNIS 297
Excluded after start
missing kalshi lock 584
stale or late market 15
no directional forecast 181
ambiguous outcome mapping 7
checkpoint timing mismatch 4