Missing data
A signal may be omitted, down-weighted, or represented as unavailable. Absence is not treated as positive evidence.
Public model card · v2026.07
How AI sports betting analysis works.
A plain-language model card for Propeller’s player-prop research: inputs, confidence semantics, update behavior, known limits, and what the system does when the evidence is weak.
Definition
Confidence means conviction—not probability.
Propeller’s confidence score describes the model system’s directional conviction after it combines sport-specific signals. A 70 confidence score does not mean a 70% chance of winning, and scores should not be treated as calibrated probabilities.
Accuracy, probability calibration, closing-line value, and profit are different measurements. The public experience does not collapse them into one claim.
Processing map
What enters the system.
MarketListed line and available pricing context
FormRecent and longer-term player performance
RoleUsage, minutes, touches, or opportunity
MatchupOpponent and environment effects
AvailabilityInjuries, lineups, and teammate effects
ContextSport-specific pace, venue, rest, and game state
Operating rules
How the product handles uncertainty.
Freshness
Current pages display the slate date or an explicit unavailable state. A prior-day row should never be relabeled as today.
Market movement
Analysis is tied to the displayed line. A changed line is a different research question and may invalidate the earlier view.
Leakage control
Forward rows are recorded when first observed in the public preview. The current API does not prove event-start or outcome timing, so neither ordering is claimed.
Known failure modes
Late scratches, sparse samples, role changes, source outages, incorrect line mappings, and unusual game states can all degrade analysis.
User responsibility
Use the score as one input. Check current lineups and prices independently, and stop when evidence is incomplete.
General AI versus a research workflow
AI can organize the question. It cannot make the source current.
| Research layer | Useful role | What still requires verification |
|---|---|---|
| General AI assistant | Structures the question, summarizes cited evidence, and presents the countercase. | Every time-sensitive fact, exact line, price, injury, lineup, and source timestamp. |
| Propeller Analyzer | Surfaces a limited public research snapshot with the displayed market context. | That the snapshot remains current and maps to the line you are considering now. |
| AI prompt builder | Creates a skeptical research protocol with fresh-source and no-action checks. | The final evidence, your interpretation, and whether the question is answerable at all. |
For a reusable checklist, start with the AI Sports Betting Prompt Builder. For a disclosed comparison protocol, see the AI player-prop benchmark.
Sport coverage
Separate models, shared disclosure.
| Sport | Representative context | Primary uncertainty |
|---|---|---|
| MLB | Pitching matchup, park, handedness, batting role | Lineups and bullpen state |
| NBA | Minutes, usage, pace, teammate availability | Late injury news and rotations |
| NFL | Opportunity, matchup, game environment | Low game count and role changes |
| NHL | Line role, shots, special teams, opponent | Goalie and line combinations |
| Soccer | Lineup, role, opponent, form | Starting XI and competition variance |
Machine-readable version: methodology-version.json. Corrections follow the editorial policy.
- v2026.07 — Initial public model card; confidence definition and forward-publication limits formalized.