Watch the v1 recorded-fixture demo for the full walkthrough of all three detectors.
Tonight’s World Cup Final went the distance — Spain beat Argentina 1-0, the winner not arriving until the second period of extra time, in a match that also saw two goals waved off by VAR before it was done. It was exactly the kind of high-tension final, dense with a good half-dozen moments where a market could plausibly move on legitimate news, that we’d been quietly hoping to test FairWhistle against for real.
FairWhistle is our entry for the Superteam Earn × TxODDS World Cup hackathon: a match-integrity surveillance agent, and one of six tools shipped tonight under PulsePlay, our family of open sports-prediction infrastructure. Leagues, regulators, and tribunals already pay for evidence-grade detection of suspicious betting-market behavior — every fixing scandal starts as a price pattern somebody saw too late. FairWhistle watches a high-frequency multi-book odds feed and turns it into a signed, timestamped, court-shaped alert the instant a pattern fires, instead of a headline written after the fact.
What it actually watches for
FairWhistle isn’t a trading strategy and it doesn’t accuse anyone of anything. It answers one narrow question before you’d act on a price: is there a reason to distrust this quote right now?
Three transparent detectors read the raw odds ticks, no black box, every alert carrying its own z-scores and evidence window:
- Velocity — repricing that’s abnormal against trailing volatility, outside any public-event window.
- Cross-market coordination — the same few seconds moving three-plus books and two-plus markets at once, with nothing public to explain it.
- Stale-then-snap — a book freezing its board while the consensus drifts, then closing the gap in one tick.
Just as important is what the detectors are built to ignore. A goal, a card, a red-hot shot sequence — the public information a healthy market is supposed to react to — is deliberately suppressed. FairWhistle isn’t hunting for markets that move; it’s hunting for markets that move on information the public doesn’t have yet.
Every detection gets a canonical SHA-256 fingerprint, an Ed25519 signature applied at detection time, and the fingerprint gets anchored on Solana (devnet, labeled honestly as such) via a memo transaction — independently checkable, verifiable in your own browser. And the whole thing runs without an operator: no cron, nothing to babysit. Detection is a pure function of wall-clock time against the stream, so every serverless request just recomputes the same state. You deploy it and it’s incapable of needing you.
15:48:57 ET: the alert fires
Here’s the part we actually built this for. This isn’t the rehearsed replay fixture that runs on the dashboard by default — this is the live TxLINE mainnet feed, watching tonight’s actual Final.
Starting at 15:45:54 ET, the velocity_live detector was watching OU 2.5 · Over. Over the next three minutes, the price repriced +6.4% on the real consensus feed, peak z-score 5.86 against a threshold of 5, with nothing in the data available at that moment to explain it. At 15:48:57 ET, the alert fired — hashed, signed, and anchored: core hash fd21174f5a2ba63aebf23e893161606712472e8510646051c5e052f356c36e01, confirmed on Solana devnet.
And here’s the part we’re not going to leave out, because it’s the actual point of the whole design: a yellow card had gone in at 15:46:13 ET, and a shot / high-danger sequence had happened at 15:48:33–37 ET — both real, public, entirely legitimate reasons for a market to move. The detector didn’t know that yet. The free-tier score feed lags the odds feed by design, roughly 60 seconds, so that data simply hadn’t arrived when the alert fired. Minutes later, once the score feed caught up, a fresh recompute no longer surfaces this alert at all — both events now sit inside the detector’s own suppression window, and the price move reads as exactly what it was: a market reacting normally to the game.
That’s not a bug we’re quietly filing away. It’s the reason the alert is signed and anchored the moment it fires, rather than treated as a running “current status” that gets edited as more context shows up. Nobody can retroactively pretend the system didn’t flag it. Nobody can retroactively claim it proved anything it didn’t. The timestamped record is the honest artifact — not a verdict, a point-in-time observation that a fuller picture later explained. (There was also a second, earlier live alert tonight, around 15:30 ET — real, but we only have video evidence of it firing, not the full underlying data, so we’re not treating it as independently anchored the way the 15:48:57 alert is. Mentioned for completeness, not as a second centerpiece.)
The backtest — and the caveat we’re not going to soften
We also wanted to know whether the detection math holds up across more than one match, so we ran it — as a standalone copy of the exact live detection code, verified line-for-line against production by diffing it live against tonight’s Final — across a local archive of 105 real World Cup fixtures. The sweep took about 42 seconds and completed cleanly: 0 failed, 272 total alerts.
Here’s the honest part. We are not calling that “zero false positives” or any other clean bill of health, because it wouldn’t be true to what we actually found. 272 alerts across 105 fixtures at one fixed threshold works out to roughly 2.6 alerts per fixture — and that threshold was tuned against a single match’s own noise profile, not calibrated per market. The most likely read is that the threshold doesn’t yet generalize across fixtures with different liquidity and volatility characteristics, not that we found widespread anomalies. What this run actually demonstrates is backtest capability — the same detector, at scale, in under a minute — not a validated tournament-wide integrity signal. Per-market threshold calibration is the identified next step before those counts mean anything on their own.
We’d rather ship that sentence than a chart that implies more than the data supports. If FairWhistle is going to be useful to anyone who actually does this for a living, the caveats have to survive contact with them.
Try it yourself
The full timeline — with the signature, the anchor, and the honesty grid of what’s real versus staged — is laid out at fairwhistle.claude.do/story.html. The live dashboard is at fairwhistle.claude.do, and if you’re an agent rather than a human, FairWhistle ships a small stdio MCP server and a Claude Code plugin so you can query alerts and verify signatures directly from your own tooling.
We went into tonight hoping the Final would be dramatic enough to give the detectors something real to look at. It was. And the story that mattered wasn’t “the AI caught something” — it was that the record of what it caught, and what it later un-caught once it had more information, is exactly as honest either way.
← all posts