Every morning at 6 AM ET, a two-agent Bayesian swarm wakes up and scans Polymarket. It doesn't care about news cycles or sentiment. It does one thing: find markets where the listed price looks wrong.
Here's exactly what it does and what you're looking at when you open the tool.
The Core Problem It Solves
Polymarket is liquid and efficient — most of the time. But efficiency breaks down in specific situations: markets with lower liquidity, events that have strong Kalshi coverage, or questions where there's a significant information gap between what the crowd is pricing and what an independent model estimates.
The scanner looks for these gaps. Specifically, it flags any market where a Bayesian swarm of two independent signals disagrees with the Polymarket price by 5 or more percentage points.
Scope note: this scanner is tuned for the liquid mid-band of the order book — contracts trading roughly between 20¢ and 80¢, where spreads are tight and a 5-point gap is a meaningful read. Sub-10¢ longshots are a different game with different math; those have their own tool covered below.
What a Mispriced Binary Market Actually Is
Polymarket and Kalshi both run on binary contracts: a contract settles at $1 if the answer is YES and $0 if it's NO, so the price — say 62¢ — is just the market's implied probability, 62%. "Mispriced" means that number is wrong: the contract is trading at 62¢ when the real probability is closer to 70%, leaving 8 points of edge on the YES side (and the mirror image on NO).
Binary contracts drift off true value for structural reasons, not magic:
- Thin liquidity. A market only a few thousand dollars deep moves on a single order. The last trade can sit stale for hours while the underlying picture changes.
- A faster book next door. When one exchange has already repriced an event and the other hasn't caught up, the lagging book is the mispriced one until the two converge.
- A model the crowd hasn't built. On questions with a clean base rate or a quantifiable signal, an independent estimate can be sharper than a lightly-traded contract's price.
The scanner exists to surface those three cases. And because the same logic runs in both directions, a binary contract that's mispriced on Polymarket is, by definition, mispriced relative to its Kalshi twin — which is exactly what the live gap table on Kalshi vs Polymarket tracks across today's games.
Two Agents, One Swarm
The engine runs two signal agents on every Polymarket question above a volume threshold.
Agent A — Kalshi Cross-Reference. Kalshi and Polymarket frequently list the same underlying question. When they do, the two prices represent two independent crowd estimates of the same event. Agent A takes the Kalshi price as a probability signal, then weights that signal by volume: the more money that's behind the Kalshi price relative to the Polymarket volume, the more weight the signal gets. A Kalshi market with ten times the volume of its Polymarket counterpart carries near-maximum confidence. A thin Kalshi market carries minimal weight but still contributes. Before any of this happens, the engine confirms the two markets actually resolve on the same event and criterion — a title match alone isn't enough, so "same name, different event" pairs are discarded rather than compared.
Agent B — Claude NLP. When there's no Kalshi match, Agent A goes silent. Agent B never does. It sends the market question to Claude (Sonnet) and asks for an independent probability estimate with a calibrated confidence weight. Claude isn't guessing randomly — it's drawing on base rates, historical context, and what it knows about the event. It's capped at 50% confidence weight, so it can't dominate the output on its own. It's a check, not a verdict.
The Bayesian Math
Each agent produces two numbers: a probability estimate and a confidence weight. Those go into a precision-weighted blend.
The market price is the prior. The formula:
posterior = (1 − weight) × prior + weight × signalIf both agents pull in the same direction — say, both think YES is 70% when Polymarket is at 55% — the swarm produces a posterior significantly above the market price. That's a +15pp edge on YES.
The swarm then runs Kelly criterion sizing: f* = (b × P_swarm − q) / b, divided by four (quarter-Kelly), capped at 5% of bankroll, with a 2% transaction cost already baked in.
What Flagged Means
A flagged market has ≥5pp divergence between P_swarm and the Polymarket price. That's the minimum threshold where the math starts to favor action.
A high-confidence flag is ≥8pp. When multiple markets hit this threshold in a given day, Pro subscribers get an alert in Discord (#premium-alerts) with the full table.
How to Use the Results
The scanner is not a buy list. It is a shortlist of markets worth investigating.
When something gets flagged, the right move is to ask: why is there a gap here? Is the Kalshi crowd seeing something Polymarket isn't? Is there a news event that landed between when Polymarket last moved and when the engine ran? Does the Claude estimate reflect something that happened after the question was posted?
Sometimes the answer is: the market is right and the signal is stale. That's fine. The scanner doesn't guarantee edge — it surfaces markets where a reasoned disagreement exists between independent pricing sources.
Combine it with your own read. The edge is the starting point, not the conclusion.
Quarter-Kelly Is the Right Default
The Kelly fraction shown is already at one-quarter of the mathematical optimal. This is intentional.
Full Kelly is mathematically optimal only if your probability estimate is exactly right. In prediction markets, your estimate is never exactly right — it's a signal with noise. Quarter-Kelly gives you roughly 75% of the EV of full Kelly while cutting variance by 93%. You stay in the game when you're wrong.
The 5% bankroll cap is an additional hard limit. No single trade — no matter how high-confidence — should risk more than 5% of what you can afford to lose.
The Engine Runs Every Morning
Results update at 6 AM ET with the daily engine run. If you check before then, you'll see yesterday's results or a pending state. The scan covers the top Polymarket markets by volume that pass the minimum activity threshold.
The Bayesian spec in plain English explains the math without formulas if you want to go deeper.
Want the tail end of the order book?
The fat-middle scan above stops at the 20¢–80¢ band by design. For sub-10¢ longshots — different math, different calibration, different exit discipline — see the companion piece: How the Live Longshot Scanner Hunts Mispriced Longshots.