Locked before, graded after, never rewritten
“Every official call is locked before the bout begins, then graded against the official result. Picks are never rewritten after the fact.” Corrections are new, dated revisions with a reason; the original stays on file. The live record and the backtest are never combined.
The same principles run through the network's pick products: a call is frozen before the event, graded from the official result, and kept. Provisional calls are not part of the record.
What to read in a record
| Field | Why it matters |
|---|---|
| Sample size | A handful of calls is noise — see Sample size |
| Hit rate | Meaningless without prices — see Break-even rate |
| ROI and its standard | UFC uses a flat 1-unit stake at the best price in the lock-time market snapshot; draws, no contests and voids are neither wins nor losses |
| Brier score vs 0.25 | Grades the probabilities, not just the picks — see Brier & log loss |
| Official vs tracking | Kept separate; never add them together |
| No-calls | Fights or games the model declined, and why |
- A record shows 7–4 (63.6%).
- If most calls were favourites around −200 (break-even 66.7%), 63.6% is below break-even.
- The ROI line, not the hit rate, tells you which world you are in.
Publication gates
Several products hold models in tracking, shadow or validation until they pass criteria set in advance — a minimum sample, a minimum number of weeks, beating baselines. A product showing "gated" or "in validation" is the process working. See Confidence & gates.
Records are backward-looking and small at first. A short record that looks great and a short record that looks terrible are both mostly variance.