Nine stages, one question.
What is the true probability of this market — independent of what the price says? Here's how we answer it. We share the architecture, not the internals.
News ingestion
Multi-source news APIs (Guardian, Reuters, AP, niche feeds) parsed continuously. Every claim is timestamped and source-tagged for downstream attribution.
Pre-trade viability
Filters out markets that are illiquid, ambiguously worded, or already resolved. Saves the model from wasting Bayesian cycles on noise.
AI sanity check
Lightweight LLM pass that catches hallucinated correlations and obvious model errors before they reach the scoring layer.
Liquidity check
Polymarket CLOB orderbook depth, slippage modelling, and minimum-volume gates. We won't signal a market you can't actually trade.
Claude correlation
News-to-market matching, trade thesis generation, and confidence scoring. Produces the human-readable 'why' for every signal.
Bayesian probability
Independent probability estimate — explicitly NOT anchored on the market's contract price. This is the core source of edge.
Whale-overlay assembly
Cross-reference the 287 tracked whale + rising wallets. If smart money is already in the market, the signal carries that context.
Delivery
REST API for the Engine bot. Signals delivered in real-time — zero artificial delay.
Outcome tracking
Every signal is recorded with predicted probability and resolved against actual outcome. Drives quarterly track-record reports.
What we don't share
The specifics of stage 06 — the priors, the feature weights, the calibration loop — are private. Publishing them would let a competitor reproduce the edge in a weekend. The architecture above is the trust signal; the math behind it is the moat.
Want signals from this pipeline?
Sharp tier — all signals, real-time, with full thesis and whale context.