Research standard
From licensed racecard to settled record.
AI is used for ranking and explanation. Application code controls identity, probability normalization, evidence binding, fair-price arithmetic, immutability, and settlement.
The analysis contract
Ranked field
Every active runner appears exactly once in model order.
Win probability
Probabilities are validated and normalized to 100%.
Fair decimal price
Calculated deterministically from the normalized probability.
Market comparison
Shown only when a permitted, fresh market price is stored.
Field-level evidence
Evidence labels resolve to non-empty source fields for that runner.
Uncertainty
Missing inputs and data completeness remain visible beside the estimate.
Audit identity
Model ID, prompt version, input hash, source time, and generation time are retained.
Settlement
Official results close the record without deleting unsuccessful predictions.
Input policy
The generation payload is limited to the normalized race, active runners, form, conditions, stored ratings, jockey and trainer fields, comments, and fresh permitted prices. Provider credentials and unrelated account data never enter the model prompt.
Validation sequence
- 1Confirm the race is upcoming, active, and has at least two runners.
- 2Require every active runner exactly once and reject unknown identities.
- 3Require total probability within a narrow tolerance, then normalize to 100%.
- 4Replace evidence text with the exact stored value for the selected source field.
- 5Calculate fair decimal odds and market gaps in application code.
- 6Retry malformed structured output once; otherwise return analysis unavailable.
- 7Store the completed prediction with its input hash and generation identity.
Interpretation limits
A fair price is a model estimate, not the unknowable true probability. Market comparison does not account for an individual's commission, limits, liquidity, tax position, staking, or jurisdiction. Performance conclusions should wait for a sufficiently large, prospectively recorded sample.