[Flotapponnier]

Spotted on the live page: OKX first to report on 95.93% of Solana trades, while its own head lag p50 there is 209 ms against Mobula's 11 ms. Two series on the same panel saying opposite things.

What the race was scoring

Any trade two feeds both reported. That was the whole eligibility rule, and it was correct for as long as every feed was subscribed to the bench pool. OKX and Birdeye are subscribed per token, so they see every pool on it while the other four see only ours. Since they shipped, the book has been dominated by trades from pools this bench does not measure, each scored as a race between those two alone.

Measured on Solana over one hour:

races closed	77,776
trades on the bench pool	~1,200
firsts awarded to OKX	75,151 (on 1,210 pool trades of its own)

And the number that makes it concrete: Mobula was first on 85 of the 109 pool trades it reported, about 78%. The page published 0.08%.

Its share is forced toward zero by arithmetic, not latency. The numerator is its wins; the denominator is every race in the window, including the tens of thousands it could not enter:

headLagFirstShare.Set(float64(wins[a]) / float64(len(hist)) * 100)

Base (84.90%) and BNB (85.90%) were the same defect, less visibly.

Why the obvious fix does not work

Voiding races the reference never saw fails on Base, where the reference is the chain-wide flashblock stream and therefore matches every transaction on the chain. Membership and timing were being answered by one refClock, and they are different questions.

So poolTrades is a second set, written only by the pool subscriptions in reference_monitor.go and never by the flashblock stream. Timing still comes from reference, so Base keeps its flashblock zero point; the new set only decides what is eligible.

Chains with no pool subscription are left unfiltered rather than voided. Robinhood has no endpoint we trust, and voiding everything there would read on the page as four dead feeds rather than an unmeasured chain.

Checking at close time is safe

The pool subscription lands within ~130 ms of the fastest feed on Solana (measured: ref lag p50 -0.106 s for Mobula, -0.130 s for Serialized, -0.040 s for Codex), comfortably inside the 3 s raceWindow. A genuine pool trade is present by the time its race closes.

race.go had no test file

Which is how this shipped. The new one asserts both directions, and I verified the discrimination rather than assuming it: with the guard removed, TestRaceVoidsTradesOutsideTheBenchPool and TestRaceMembershipDoesNotUseTheChainWideReference fail, while the two tests guarding against over-voiding still pass.

Expected effect on published numbers

Solana drops from 77,776 races an hour to roughly 1,200, OKX's first share falls to whatever it earns on the bench pool, and Mobula's rises toward the ~78% it was already achieving. The Solana head_lag column moves with it, since that figure is produced in closeLocked.

Base's head_lag does not move: it comes from resolveOne against the chain-wide reference, which is the separate scope question still open.

Gates

go build ./..., the full go test ./cmd/script/ suite, and pnpm validate (235 specs valid) all pass. Methodology now states the eligibility rule and says plainly that figures published for OKX and Birdeye on the Solana column before today were measured on a wider set of trades than the other four feeds.

🤖 Generated with Claude Code