How TapeXray measures Kalshi vs Polymarket gaps (2026)

TapeXray counts a Kalshi-Polymarket pair as one bet when both mids are between 15¢ and 85¢, the gap is at most 12¢ and under 35 % of the cheaper mid, confidence is at least 0.75, and both sides traded $1,000 in 24 h. Fees are priced twice: at Polymarket's 7 % maximum, then at each market's rate.

Where do TapeXray's numbers come from?

Everything on this site is computed from one archive that TapeXray collects itself: the public market catalogs and order books of Kalshi and Polymarket, read on a schedule and stored as dated partitions. Nothing is typed in from a venue's marketing page. The archive currently holds 180,476 markets, 9,826 of them on Kalshi and 170,650 on Polymarket, and its newest catalog partition is dated 2026-11-17. The market pages show only markets that cleared $100,000 in traded volume, because a page about a market nobody trades would be a page about a quote, not a price.

The live board is built from that same archive by a matcher that pairs Kalshi and Polymarket sub-markets on their title tokens and then reads the latest order book on each side. A pair enters the board with two mids, a raw gap, the matcher's confidence score and both sides' 24-hour volume; the gate described in the next section decides whether it is shown as a real pair, as a pair the fees eat, or disclosed as suspect with the reason. Every figure on this page is read from that export when the site is built, not from the prose around it, which is why the counts here match the board's own on the same deploy.

Which matched pairs count as the same bet?

A title match is not proof. The matcher cannot tell one team's win market from the other's, and two venues' longshots are routinely two different questions, so the board applies a same-bet gate to every matched pair before it prices anything. The gate is version 4; its rules and thresholds are read from the code that applies them, so the table below cannot describe a gate the board is not running. A pair that fails any rule is not deleted: it is exported as suspect with the rule that rejected it, and the board discloses the count.

The same-bet gate, version 4
RuleThresholdWhy it exists
Both mids inside the band15¢ to 85¢Longshot books quote in whole cents, so a small gap on a cheap side looks large; two venues' longshots are routinely two different bets.
Gap relative to the cheaper midat most 35 %A same-outcome gap is a few percent of the price, not a multiple of it.
Raw gap between the two midsat most 12¢Same-outcome prices on two venues rarely differ by more than a few cents.
Title-match confidenceat least 0.75The matcher pairs sub-markets by title tokens and cannot tell one team from the other on its own.
Traded volume, both sides, last 24 hat least $1,000Under that a mid is one resting order wide of nothing; a missing volume fails the floor rather than passing by default.
Source: pmsite.xmatch's gate constants (suspect_reason), read at export time

The funnel below is the last scan's, counted over the full lists rather than the capped tables the board displays. Most of the loss is at the top, in books that were dead or missing and in pairs the gate rejected; the number that survives to "positive after fees" is small on most days, and that is the honest shape of the data, not a failure of the scan.

The last scan's funnel, 2026-11-17 06:05 UTC
StagePairs
Matched pairs from the matcher8,153
Dropped: no order book on one side0
Dropped: dead book (a side pinned at the edge)2,805
Dropped: stale book209
Rejected by the same-bet gate (disclosed as suspect)5,132
Passed the gate7
Passed, but round-trip fees eat the gap7
Positive after fees0
Source: xmatch.json's funnel block, gate version 4, counted over the uncapped lists

How are fees modelled, and why in two passes?

Both venues charge takers on expected earnings, rate x P x (1 - P) per contract, where P is the contract price in dollars, so the fee peaks at a price of one half and shrinks toward either end of the range. Kalshi's rate is one number, its general schedule of 0.07, rounded up to the next cent on an order. Polymarket has no single rate: each market publishes its own on its catalog row, by category, from fee-free to a maximum of 0.07. TapeXray measures that distribution instead of typing it, from a sample of 20,000 of the 291,902 catalog rows the archive holds; the Polymarket fee page publishes it with row counts.

Pricing every one of the matcher's pairs at its market's real rate would mean a catalog read per pair, which exceeds the export's memory budget. So the board prices in two passes. Pass one gates and ranks every pair at the Polymarket maximum, the conservative direction, which means no pair is ever admitted that its true rate would have excluded. Pass two re-prices only the pairs that passed the gate at each market's own catalog rate, so the fee-adjusted gap can only rise on re-pricing. A market with no readable schedule keeps the maximum and says so in its fees block. The fee calculator runs the same formula on both venues, and the Kalshi fee page works through the rounding.

The two fee passes
PassPolymarket rateKalshi rateApplied to
Pass 1: gate and rank0.07, the published maximum0.07, general scheduleevery matched pair (8,153 on the last scan)
Pass 2: re-priceeach market's own catalog rate (10 measured categories)0.07, general schedulegated pairs only (7 on the last scan)
Source: pmsite.fees' rate constants and pmsite.xmatch.FEE_NOTE / reprice_gated

What do per $100 and per $500 mean, and why does each surface keep its own unit?

Two conventions appear on this site and they are deliberately not mixed in one view. The board's gaps are in cents after round-trip fees: the difference between the two venues' mids, minus taker fees on both legs, doubled for entry and exit, because a bettor who crosses a gap between two mids has to get out again. That is the number the board ranks by and the number the "fees eat it" list is measured against, and it can only be compared with other round-trip figures.

A fork is a different trade. Buying YES on one venue and NO on the other for a total under $1 holds to resolution, so there is no exit leg and each side pays one-way taker fees once: Kalshi's rounded up to the whole cent, Polymarket's at the market's own catalog rate. The fork card and the alert bot express that as money on a $500 payout, "put in this, get back $500, keep the difference", scaled from the per-pair figure so that a reader sees a stake they can picture rather than a margin in fractions of a cent - per $100 the same lock would read as one to three dollars and undersell it. When the depth of the books is not known, the card says so, because the book decides how many pairs it takes to fill $500. Neither convention is converted into the other on a page, and a figure quoted from this site should say which one it is.

How fresh is the number on the page?

The site is a static export of the archive, rebuilt on a schedule every 6 hours. Each rebuild re-reads the catalogs, re-runs the matcher's gate over the latest order books, re-prices the survivors, and writes the numbers on every page from that one export, so a figure is never fresher than the last deploy and never older than it either. The board's own stamp is the time its pairs were scored, 2026-11-17 06:05 UTC; the fee table's stamp is the date its catalog sample was drawn, 2026-11-17; the archive's stamp is the newest catalog partition, 2026-11-17. The three can differ by hours and that difference is real: a book is read more often than a catalog is re-listed.

The gate version travels with every recorded day, and the rolling windows on the board restart whenever it changes, so a figure from one gate regime is never averaged into another. The availability changelog records what changed on the state pages and when; it is not this page, and it does not describe method.

What does this site not verify?

A methodology that lists only what it checks is a brochure. The table below is the standing list of claims this site relies on but has not independently verified, with the reason. Each row exists as a documented fact in the project's own records before it appears here, and the status column reads the export's own flags where one exists, so "not re-verified" turns into "verified" on the deploy after the check succeeds and not before. If a row here is wrong, or a claim on the site is missing from it, the contact page reaches the desk directly.

What this site does not verify
ClaimStatusWhy
Kalshi's 7 % general taker ratedocumented, not re-verifiedkalshi.com's fee-schedule PDF has returned HTTP 429 to every fetch since 2026-09-21; the rate is retained from the previously read schedule.
Kalshi per-series fee multipliersunpriced; every series is charged at a multiplier of 1.0Kalshi's series endpoint carries a fee_multiplier, but the archive has no series table, so a series above 1.0 is priced too generously here.
Polymarket's per-category rate tablemeasured from a sample of 20,000 of 291,902 catalog rows (7 %)Reading every row's schedule exceeds the export's memory cap, so a rare category can be under-represented in one day's read.
That a gated pair is the same outcomegated, not provenThe matcher works from title tokens; the gate rejects the implausible pairs and discloses them as suspect, but it cannot read a rulebook.
Kalshi's close_time as a settlement datenot used as oneIt is a trading ceiling that can sit days after the event, so the two venues' settlement timing is disclosed as a risk rather than gated on.
Source: docs/forks.md §11, docs/measurements/2026-09-23-fee-sample-oom.md, fees.json's verified flag and pmsite.xmatch's gate comments

What does the weekly counter add up, and how is it graded?

The line under the live board's heading adds up what the alert bot found this week, priced at the moment of the alert. A gap or a lag is valued by walking the cheaper venue's ask ladder with a $100 stake: its edge is the pricier venue's mid minus the average ask paid on the cheaper venue, per contract, less that venue's one-way taker fee, and it is labelled "if the pricier venue is right", because the gain only exists if the other venue's price was the correct one. A fork is valued by the money the depth walk keeps after both venues' fees, for at most a hundred pairs (a deeper walk is scaled down to a hundred), and labelled "locked in either way" - a different resolution source or cut-off can lose a fork both legs; read both venues' rule pages before you bet. The week starts on Monday at midnight UTC; each market and side counts once a week at its best value among this week's alerts, repeats are listed but never summed, one fork is one market and direction however often it was re-alerted, and a fork on a market suppresses that market's gap or lag while both are open. Only alerts that passed the public gates count; a gap or lag whose ladder could not be read counts as an opportunity worth nothing, a fork whose depth could not be read is not counted, and an alert whose walk found no edge left is not counted at all. This week's total is $0.00 across 0 opportunities, counted since 2026-09-30. Membership guarantee N: 12, four times the lowest full week, never under one.

This week's counter, from 2026-11-16
KindCounted this weekValueLabel
Forks0$0.00locked in either way
Gaps0$0.00if the pricier venue is right
Lags0$0.00if the pricier venue is right
Whale moves——followed at alert time; counted once the whale record grades them
Multi-venue arb——locked in either way; counted once that vertical ships
Source: xmatch.json's counter block, built by pmsite.counter from the case archive at export time

The record below is cumulative and graded, never projected. A gap is graded twice: once at close, as the cheaper venue's bid when the case ended minus the cost paid, less both one-way fees; and once at resolution, from Kalshi's settled result, which grades the Polymarket leg too because a matched pair is one bet. Each average is over graded cases only and prints its count beside it; pending cases are listed, never averaged. A fork is graded by whether both venues resolved the same way: the count is over forks graded on both venues, a fork still waiting for one leg is listed as pending, and a Polymarket result is read as Yes or No only once the catalog says which outcome is which.

The graded record, cumulative
RecordGradedPendingPer $100 staked
Gaps: bought at the alert, sold at the close0——
Gaps: held to resolution00—
Forks: resolved the same way on both venues5 of 5106—
Source: xmatch.json's counter block, built by pmsite.counter from the case archive at export time

What does capturable mean on the Edge Monitor?

The Edge Monitor's quiet-day line - "N cases, +$X capturable" - is computed at export time from the monitor's own gap ledger, under one conservative rule (methodology version 3). A case counts when it has closed inside the rolling 30-day window, passes the week board's gate-compatibility admission, and is worth at least $1 on its own stake. Each case is counted on that stake: $100 a gap, $500 a fork. A case's measured size is the depth a full scan's book walk recorded at its peak, else the live walk's on its latest beat, else none. A case never walked counts nothing (88 of the window's 498 cases right now), nor does a walk that found no depth. Per $100, a gap contributes the walk's average edge per contract times the contracts $100 buys on the cheaper venue's peak mid; a fork, the walk's net per pair times the pairs $100 buys at the peak legs, counted five times. Neither passes the walked depth; a larger stake earns more only up to the book's dollars. The sum, rounded to cents, is the capturable figure.

Capturable is what the tape showed, not what a holder earned: an opportunity, not a promise. Depth past the walk's last readable ladder, a leg that resolves against the position, and fees the walk did not price all come out of a real fill, and none of them are in the number. Right now the rolling 30-day window holds 498 closed cases, capturable at +$330.27 with $100 on each gap and $500 on each fork, over — days the scanner actually ran.

The site's headline adds one more line to the capturable figure above: the whale bets, $100 a day, spread evenly over the bets a whale put $100,000 or more on that settled that day, each at the whale's own entry price, counted as what came back less the stake, with no fees or slippage in it. Over the window it stands at −$82.28 ± $56.39 across 1,612 bets; the ± is the line's day-to-day swing, so a figure inside it is noise, not edge.

The monitor's compound tiles replay the same per-case profits as a running stake, one tile per kind in its own unit: $100 rides through the window's won gaps, $500 through the won forks, each won case reinvesting its peak profit per unit. A gap's position is YES on the cheaper venue, and it won when that venue's leg resolved Yes (a Polymarket "split" pays half and is not a win); a fork won when both legs settled the same way. A pending case (a leg still open) and a lost case compound nothing, so each tile names the won count beside the graded count and never reads as the whole window's record. Right now: 46 won of 240 graded closed gaps, $100 compounded to $134.88; 1 won of 1 graded closed forks, $500 compounded to $506.19.

What would the free channel have made over the same window?

The free channel posts 6 hours after the alert goes to members, so the capturable figure above is the members' side of that delay. The free side is built from the same cases and the same whale bets, at the same stake, each entered at the price the market showed 6 hours after the first sighting; the free channel's reader would have come to +$14.59 on the gaps and forks and +$45.54 on the whale line over the same 30 days. A gap, fork or whale bet that was already over by then counts as nothing. A gap or fork whose price at that moment was not recorded is counted at its peak, which can only understate the members' lead; a whale bet with no price then counts as nothing. Neither figure carries fees or slippage. The whale figure is quoted only when the export could price every late whale entry, so a failed lookup shows up here as "not priced", never as a smaller figure.

How long do detected gaps and forks live?

The Edge Monitor labels every open case with how long similar cases lived, and the table below is where that label comes from. At every export the lifetimes of the alert bot's own cases are measured from the case archive, the append-only record both alert lanes write and the one source the ledger's month-long retention cannot lose. A case's lifetime runs from its opened event to its closed event, and only closed cases have one: a case still open at export time is excluded, and the exclusion is counted (0 open right now), because mixing still-alive cases into "lived" would understate every figure. A case whose opening event fell off the archive's read window is excluded too - its start is not honestly known. The record begins 2026-09-30.

How long detected cases lived, closed episodes since 2026-09-30
KindClosed episodesMedianMiddle halfShortest to longest
Forks1101 hour59 minutes to 2 hours58 minutes to 2 days
Gaps8823 minutes1 minute to 8 minutesunder a minute to 41 hours
Lags0———
Source: gap-stats.json, built by pmsite.gap_stats from the case archive at export time

Read the gap row as: of 882 closed episodes, half lived between 1 minute and 8 minutes and the median was 3 minutes; the fork row reads the same way. The quartiles are linear-interpolated between the nearest ranks, and the count beside each kind is the whole claim - when n is small the range is a handful of episodes, not a law, and the cards print it so the reader can judge. These are the lifetimes of episodes the bot detected and watched close, not of every spread that ever existed; lags are alerted without a ledger case, so their row reads zero until that changes. This is methodology version 1, and the live board's cards carry the same numbers.

By The TapeXray deskCopy last reviewed

See the live board this page describes →

Gaps close in hours.

The free channel posts at the next rebuild; members' DMs arrive as the scan finds it.