How PROVISUM works

1 · The data pipeline

PROVISUM runs a recurring crawl — by default every 10 minutes — across five venues:

• Polymarket via the Gamma API for market metadata and the CLOB endpoint for real order-book depth.
• Kalshi via the trade-api v2 events feed (we walk events, not the parlay-flooded flat market list).
• PredictIt via its public market-data feed (top-of-book).
• Smarkets via its v3 events → markets → quotes chain.
• Limitless — an on-chain CLOB on Base — where we fetch each durable market's real order book (not the mid) so the YES/NO asks are genuinely tradeable.

Every crawl is live. Nothing on the platform is mocked or simulated. Dead books (no offers), zero-volume markets and auto-generated parlay combinations are dropped before matching. Each signal is written to a database with a first-seen timestamp, which is what powers the "detected 6m ago" ages, the sparklines, and the recently-closed feed.

2 · Matching the same event

Two markets are only an arbitrage if they resolve on the same outcome. We match with IDF-weighted token similarity: generic words ("election", "2026", "will") are common across thousands of markets and prove nothing, so rare tokens — candidate names, team names, strike prices — carry the weight. On top of that sit hard guards:

• Strike guard — "above $2,000" never matches "$980 or above".
• Date guard — "by September" is not "in December"; markets whose deadlines are far apart are rejected.
• Parent-question guard — a shared rare name isn't enough; the questions themselves must overlap ("LA Mayor winner" ≠ "who Bernie endorses").

3 · Pricing the hedge (with real fees)

For each matched pair we compute both directions — YES here + NO there, and the reverse — and keep the better one. Crucially, every venue's fee model is baked in:

• Kalshi: taker fee of ⌈0.07 · P · (1−P)⌉ per contract at entry.
• PredictIt: 10% of profit on the winning side at settlement.
• Smarkets: 2% commission on net winnings.
• Polymarket: modelled fee-free unless a market flags otherwise.

So the spread you see is the spread modeled net of fees (before slippage) — not a gross number that evaporates at settlement.

4 · Verifying executable depth

A spread at the top of the book is only real for the size sitting there. One click walks the actual order books (Polymarket CLOB, Kalshi order book) for your contract count and returns the volume-weighted average price, the depth available, and whether the edge still clears after you've eaten through the book. It routinely shows that a tempting 9% top-of-book edge only exists for a handful of contracts.

5 · Why our signals aren't junk

The hardest part of arbitrage detection isn't the math — it's not fooling yourself. PROVISUM actively excludes false positives:

• Cross-platform matches below ~75% confidence with a large "edge" are flagged suspect and hidden — they're almost always scope mismatches (a single district paired with control of a whole chamber).
• Same-platform ("cross-market") arbs are only emitted for events verified to be genuinely two-outcome and mutually exclusive — two candidates out of a five-way race are not a hedge, and we reject them.

6 · Data honesty

We would rather show you fewer, real opportunities than a wall of impressive-looking fakes. Every row carries a match-confidence figure and a verification button, because the final check — do these two markets truly resolve on identical rules? — is one only you can make. Prices are pulled on an interval and can differ from what a venue shows right now; treat the terminal as a scanner that points you at where to look, not as a quote you can trade on blind.