The scan takes the best Over and the best Under across books at one exact number. It is symmetric: no book is preferred, and the two legs are chosen by price alone.
Legs at exchange and prediction market venues are kept out of pairing entirely. Their prices carry commission or per-trade fees, plus spread and depth, which the posted number does not state -- so a margin computed as though those legs were free would be a false claim on the one screen that invites staking both sides. Their prices still render on the board and in line movement; only the margin claim needs the netting.
A book we have never classified stays IN. The exclusion runs on a known venue role, never on a guess about an unfamiliar name.
Decimal prices above 99 are excluded, because a price that long is indistinguishable from an American one to the shared detector and would be reinterpreted as a bet nobody offered.
A fat margin is not a better opportunity. It is usually evidence that one book is mispriced against its peers, which is the price most likely to be voided, re-graded or pulled before the second leg lands -- and the pattern a retail risk desk acts on hardest.
So each leg is checked against the other books quoting that same number, and a leg sitting 3 percentage points or more of de-vigged probability away from those peers is flagged. The default ordering is outlier_last: flagged rows are demoted rather than clean rows promoted, because there is no reason to rank one clean row above another beyond the margin it already shows.
The peer set deliberately excludes both legs of the opportunity. Two books are needed for an arbitrage to exist, so a consensus that included them would be the two legs comparing against themselves.
An unchecked leg is not demoted. Ranking a leg nobody could check alongside one that was checked and found out of line would state a finding we do not have.
A row whose number was quoted by nobody except its own two legs carries no verdict in either direction, and says so in the payload rather than reading as a clean row:
Each leg reports how many minutes its book had been sitting on that price when the capture was taken. That is not the age of the snapshot, which is the same number for every row: the two legs of one opportunity can differ by a long way, and a book that last moved forty minutes ago is a different execution risk from one that moved two minutes ago in the same snapshot.
The two ages add. Total staleness is the leg's own figure plus the snapshot's, and both are served separately so the sum can be done and checked.
A price stamped after our own capture is not an age but a clock we cannot reconcile, so it reports unknown rather than zero.
This tool is part of the Market tier at $4.99/month. Goes live in October.
This page describes how the tool works. It carries no price, no fair-value figure and no margin: every number on it is a fixed threshold or a count read out of the product's own configuration.