Methodology

All market data comes from Polymarket's public Gamma, CLOB, and Data APIs. Probable never places, cancels, or manages orders, and never requests or stores a Polymarket private key or API credential.

1. Order-book midpoint, when both a best bid and best ask exist.
2. Recent last trade (within 10 minutes), when the book is empty or one-sided.
3. Gamma's outcome price snapshot, when no live CLOB data is available.
4. INSUFFICIENT_MARKET_DATA — returned explicitly rather than inventing a price.

Every non-midpoint price carries an explicit warning naming which fallback was used.

The displayed probability is the market's current price — what traders are willing to pay, not a statement of certainty. Signal confidence is a separate heuristic measuring how trustworthy that price is, given liquidity, related-market agreement, and resolution-rule clarity. It is never a statistical confidence interval and never a prediction of the real-world outcome.

signal_confidence = 0.60 × market_quality + 0.25 × related_market_agreement + 0.15 × resolution_clarity, where resolution_clarity = 100 − resolution_risk. Grades: 80–100 HIGH, 60–79 MODERATE, 40–59 LOW, 0–39 VERY_LOW.

A move from 40% to 48% is reported as +8 percentage points, never "+8%" (which would mean 40% → 43.2%). All movement figures in Probable use percentage points.

0.25 × spread_score + 0.25 × depth_score + 0.20 × activity_score + 0.15 × open_interest_score + 0.10 × freshness_score + 0.05 × concentration_score

Every component is returned alongside the total — a score without its components is considered incomplete. Depth, activity, and open-interest scores use a log scale so that early liquidity counts more than later liquidity at the margin. Top-holder concentration is approximated from Polymarket's visible top holders, not true total supply (which the public API doesn't expose).

An LLM extracts findings and their exact evidence spans only — it never assigns the score.

• Missing resolution source: +25
• Ambiguous or conflicting deadline: +20
• Timezone omission: +5
• Undefined decisive term: +15
• Subjective decisive verb: +15
• Question/description conflict: +20
• Overlapping or non-exhaustive outcomes: +15
• Missing edge case: +10
• Other finding: +3 / +7 / +12 (low / medium / high)

Capped at 100. Bands: 0–19 LOW, 20–39 MEDIUM, 40–69 HIGH, 70–100 CRITICAL.

A finding is dropped — never shown — if its cited evidence isn't an exact, locatable substring of the market's actual text.

Multi-outcome sum(mutually-exclusive outcome groups): flagged only when the combined price exceeds 1.0 by more than a 3-percentage-point base buffer plus half of each market's visible spread — never on a bare excess above zero.

Near-duplicate: two markets with near-identical wording and the exact same stated deadline, priced more than the requested minimum edge (default 3 pp) apart. This is a lexical-similarity check, not a verified semantic equivalence — every result is labeled a candidate inconsistency requiring manual verification, never "arbitrage" or "certain mispricing."

A third mode, logical implication between related markets, requires an LLM relation classifier and is not yet implemented — it is accepted as a request parameter but produces no candidates, with an explicit warning rather than a silent no-op.

• Search relevance currently uses lexical token overlap, not an LLM-based semantic reranker.
• The last-trade fallback (hierarchy step 2) isn't wired up yet; pricing falls straight from the order book to Gamma's snapshot price.
• No caching layer yet — every response is currently a live read.
• Holder concentration is a proxy among visible top holders, not total supply.