1. Jevan makes separate predictions for each market.
Jevan evaluates model forecasts, expert research, and game context. An ATS forecast asks which side covers a verified reference spread. A moneyline forecast separately estimates an outright home win, away win, or tie, including overtime. An Over/Under forecast estimates whether both teams’ combined final points finish above or below a verified total, including overtime; equality is a push. Cover, win, and totals probabilities answer different questions.
Our objective is the most likely outcome: which team wins outright, which side covers the published spread, or whether combined points finish Over or Under the published total. This supports straight-up and spread-based office pools. We do not optimize selections for betting payout or expected profit. A pool using a different spread is asking a different question.
Before each new prediction, an AI evidence reviewer checks a separate packet for that game and market. It selects relevant original reporting, preserves the source forecasts and statistical data, and can research missing facts with dated citations. Its role is editing and research; Jevan makes the pick. Missing or unapproved evidence holds the prediction for review.
For new reviews, Jevan evaluates the same approved game-and-market prompt 20 times. We average probabilities for the same outcomes, then select the more likely side. The ChatGPT evidence review runs once for that shared prompt. These repeated estimates are not 20 independent sources or a guarantee of greater accuracy. An incomplete batch leaves the previous published pick in place.
After Jevan decides, a separate AI explanation step receives the exact saved prompt and decision. It writes three short reasons supported by that evidence. These are labeled AI evidence summaries, not Jevan’s internal reasoning. The explanation step cannot change the selection or its probability, and does not look up later news or results.
ATS cover and Over/Under probabilities exclude a push. Moneyline probabilities include ties as a separate outcome. None is a measured success rate or a guarantee. Picks are ranked within their own market, not by comparing different kinds of probability.
Every ATS selection applies to the exact published reference spread, which may differ from a sportsbook’s current line. Moneyline selections require captured odds and an eligible outright-win forecast. Over/Under selections require an exact total, captured selected-side odds, and an eligible totals forecast. Missing inputs, an unavailable engine, or a forecast received after kickoff cannot produce a new published pick.
2. One free pick in each market, chosen before the result.
One highest-ranked eligible selection from each weekly market becomes free: one ATS pick, one moneyline pick, and one Over/Under pick for one game. Their sides, spreads, totals, odds, and publication times are recorded before kickoff. The free game stays the same. Picks begin as provisional, with scheduled game-specific reviews and late-news checks. Accepted revisions retain their own publication time, side, reference line, odds, and confidence in the history. Picks lock 10 minutes before kickoff. Only the final version is graded, including losses, and it cannot change afterward. A failed or missed review leaves the last completed forecast visible with a warning. Each market initially publishes only when its required inputs and forecasts are ready.
When we review each game.
- Initial review: ATS, moneyline, and Over/Under picks publish when the weekly schedule, required inputs, and forecasts are ready.
- 24 hours before kickoff: a fresh review of each game and market using updated research and game context.
- 90 minutes before kickoff: another review using the latest available information.
- Late-news checks: about every ten minutes during the final 90 minutes. Changed evidence can trigger an additional review.
- 10 minutes before kickoff: picks lock. Only the final published version counts toward the track record.
Review times are target start times for each game, not a promise of immediate publication. Queue and provider delays can affect completion. A failed or missed review retains the last completed pick with a warning; late results cannot change a locked pick.
3. Self-learning through outcome feedback.
Jevan’s learning loop compares its own saved pregame forecasts with verified final scores. It automatically fits a confidence-band reliability memory and saves a new version when the evidence changes. After ten eligible decisions, learned reliability and balanced past examples inform future analysis. Only results known before the next prediction can inform it, and ATS, moneyline, and Over/Under evidence remain separate.
The system evaluates a learned calibration estimate alongside raw predictions on later observations. That shadow evaluation does not override Jevan’s published probabilities or prove that revised prompts improve its future picks. Accuracy and calibration must be measured on new games.
This updates the evidence Jevan considers. It does not retrain the underlying AI’s weights, promise a better win rate, or rewrite previous picks. While the sample is still small, the system collects evidence.
Current learning status: Collecting verified Jevan outcomes for dated feedback to later forecasts. Legacy aggregate results do not train Jevan.
4. Results update after the final.
Automatic checks grade saved selections against verified final scores, including overtime. ATS grading adds the recorded spread to the selected team’s score; a positive adjusted margin wins, zero pushes, and a negative margin loses. Moneyline grading uses the outright winner; a tied final refunds the stake. Over/Under grading compares the combined final score with the original total: the selected side wins only when points are above it for Over or below it for Under; equality pushes.
In-progress games stay pending. Pushes and moneyline tie refunds remain visible and are excluded from win percentage. If a results feed fails, previously verified results are retained. Missing or invalid original selections stay ungraded.
5. Moneyline and Over/Under returns use the original odds.
The moneyline and Over/Under records risk one unit on each published pick. A win at +150 earns 1.50 units; a win at −200 earns 0.50 units. A loss costs one unit, and a push or tie returns the stake for zero net units. These examples explain the arithmetic and are not picks.
ROI is net units divided by stakes on graded picks with valid original odds. Pending and unpriced selections do not count toward settled returns. We do not invent a standard payout for ATS picks when no payout odds were recorded. Prices are historical references, not live offers or a promise that a sportsbook will offer the same terms.
6. Historical patterns add context.
The weird stat finder explores historical ATS situations with sample sizes and uncertainty alongside each record. Discovery seasons through 2021 are separated from a later retrospective holdout. Patterns that do not hold up remain exploratory.
Those records use historical market spreads, not our published reference lines. A historical pattern does not establish a future edge.
7. Discuss the board in your AI app.
Members can connect a compatible AI app using our MCP server, sign in, and approve read-only access to the published board. The connection provides picks, recorded confidence, original lines and odds, and grading context. Your AI app’s follow-up interpretation is its own analysis, not an additional Jevan prediction.
Connections can be revoked in AI connections. Private research inputs, payment information, and admin tools are never shared. Members cannot alter picks or results.
Historical data attribution.
Historical schedule research is derived from nflverse / nflfastR schedule data and work by Lee Sharpe. See the schedule data release and CC BY 4.0 license.