Posts
The Second Order
// what happens once every team's model looks the same
What Post II described — a model quietly predicting an outcome, recalculating as new data arrives — has a well-documented failure mode in the one industry that adopted algorithmic prediction decades before cycling did. Finance calls it spoofing: placing signals into a market specifically to mislead another automated system or trader into misreading what's actually happening, then acting on the false read. It's been an explicit federal crime since 2010, not a gray area — the Commodity Exchange Act was amended by Dodd-Frank specifically to name it and prohibit it.
The prosecutions since then haven't been symbolic. Navinder Sarao's spoofing was tied directly to the 2010 Flash Crash and cost him $38.6 million. JPMorgan's 2020 settlement, at $920.2 million, remains the largest penalty the CFTC has ever imposed — for the same underlying maneuver, at institutional scale. Regulators built an entire enforcement apparatus around one idea: that feeding a rival's system false signal is powerful enough to be worth a decade of prosecution.
Game theory gives the same idea a name from the opposite direction: once opponents are running similarly strong strategies, the edge stops being "play the objectively correct move" and becomes "play the move that exploits what your specific opponent's strategy assumes about you." That's precisely the arc poker AI took. Libratus beat four elite human professionals over 120,000 hands using a strategy grounded in game-theoretic equilibrium play. Pluribus went further, beating top professionals in six-player no-limit hold'em — a genuinely harder problem, since equilibrium concepts that work cleanly in two-player games don't translate directly to multiplayer ones, which pushed the underlying approach toward exploiting specific tendencies rather than relying on a single fixed optimal strategy.
Here is where this series stops reporting and starts arguing. Post II established that predictive modeling of breakaway outcomes is real, public, and technically unremarkable — the kind of tool that diffuses easily once it exists. If most WorldTour teams eventually converge on similarly capable versions of that tool, using largely the same public race data, the logical next competitive edge isn't a better model. It's exploiting what a rival's model — or a rival director sportif reading the same signals a model would flag — is likely to conclude, and engineering a situation that leads it to the wrong conclusion on purpose. Finance has a decade of case law describing exactly this maneuver. Cycling, as far as we can find, has never had to write a rule against it, because nobody's had a public, prosecutable reason to yet.
Once you can model what your opponent believes about you, the game stops being about your own hand.
— framing drawn from published poker-AI research, paraphrasedWe have found no evidence that any WorldTour team has done this, attempted it, or is even thinking about it in these terms. That's not a hedge — it's the honest state of the record. Finance and poker both took years of increasingly capable, increasingly similar competing systems before deliberate exploitation became the dominant strategy rather than a curiosity. Cycling's predictive modeling, per Post II, has been running quietly since 2019 without producing a documented version of this maneuver yet. Whether it ever does may depend on something this series can't forecast: whether enough teams converge on similar tools to make exploiting the convergence worth the risk.
Dodd-Frank's spoofing prohibition, the Sarao and JPMorgan penalties, and the Libratus/Pluribus results are all independently documented, dated, and sourced below. Everything connecting those facts to professional cycling — the entire argument of this post — is our forecast, not a documented finding. No source in this post or the ones before it describes a cycling team engaging in anything resembling model-deception. We think the pattern is worth naming before it happens, not after.
Dodd-Frank's 2010 amendment to the Commodity Exchange Act defining and prohibiting spoofing, Navinder Sarao's guilty plea and $38.6 million penalty tied to the 2010 Flash Crash, and JPMorgan's 2020 settlement of $920.2 million are drawn from CFTC and Department of Justice enforcement reporting, treated as Tier 1. Libratus's 2017 result against four professionals over 120,000 hands and Pluribus's 2019 result in six-player no-limit hold'em, including its Science cover placement, are drawn from peer-reviewed publication in Science and associated Carnegie Mellon University reporting, treated as Tier 1. No source describing cycling-specific model deception exists in this post, because none was found; that absence is stated directly above rather than implied.
If a team ever tried this, it wouldn't need to hack a rival's telemetry. It would just need a camera. Post IV, The Camera's Pulse, is the series' other forecast post — and the harder of the two to justify.
