Methodology & limitations
The product is a paid, evidence-backed decision oracle — not an LLM guessing probabilities, not a custodial betting service, not a guaranteed-win product.
The Deterministic-Core Law
- Deterministic code produces every probability.
- An LLM may route and narrate, but may not create, alter, veto, or sanity-check a probability.
- Unsupported or ambiguous markets fail closed.
- Unknown entities never silently become probability zero.
- Every result identifies its model version.
- Every successful or refused decision remains auditable.
- Model agreement is not automatically empirical calibration confidence.
- Missing data stays missing, not zero.
- The evidence and receipt ledgers are append-only.
How a probability is produced
A market is first parsed into a canonical form: the verbatim resolution rule, the named settlement source, the resolution time, the bound entity/location, the strike, and which side YES prices. If any of these cannot be bound, the oracle refuses. Otherwise the market is routed to the matching deterministic engine, which computes an independent probability from real-world sources with an explicit uncertainty interval and a versioned model. No LLM creates, alters, vetoes, or sanity-checks that number.
Edge, cost, and restraint
The edge is oracle_probability − implied_probability, where implied comes from the executable
bid/ask midpoint, not the last trade. An edge is only called actionable if it clears both the
oracle's own uncertainty band and real trading friction:
- Kalshi fees use the published taker formula (0.07 · P · (1−P)).
- Limitless fees use the conservative upper bound of the official published buy-fee table.
- Polymarket fees are not yet verified — friction there uses the real spread only, and the missing fee term is disclosed, never guessed. A market whose fee cannot be quantified is refused, not over-called.
Confidence is not calibration
Confidence measures how tightly the deterministic ensemble agrees right now. Calibration is the measured, precommitted historical hit rate with its independent sample count. The site never presents model agreement as proof of accuracy, and never shows a hit rate without the number of independent events behind it.
Limitations
- No forecast is guaranteed to win. Probabilities are estimates.
- Coverage is deliberately narrow — data-resolvable markets with a wired engine and a verified source.
- Markets whose resolution rule, source, time, entity, strike, or YES side cannot be bound.
- Venues whose fee schedule is not yet verified (edge is disclosed, never guessed).
- Anything requiring subjective judgement rather than a data source.
- Cross-venue edges are never risk-free: fill, custody, venue, cancellation, and settlement risks remain.
- Funded execution is disabled; this phase places no orders.