Hook: In a recent F1 digest, two facts caught my eye that I initially read as separate stories, but then realized they're part of the same engineering play. First — Audi chief Mattia Binotto openly stated that "the ADUO system is being exploited against its original intent": a team with an overall advantage can afford not to reveal the full potential of its engine to artificially qualify for additional upgrade allocations. That's exactly how, in his view, Mercedes could have operated — winning 7 out of 9 races while still getting PU upgrade rights. Second — an academic preprint on arXiv (Kleisarchaki, 2603.01290v3) that formalizes the 2026 regulations as a Partially Observable Stochastic Game and describes a phenomenon the author calls a "counter-harvest trap" — a strategy where a car deliberately conceals energy deployment signals to deceive its opponent. And here's where it hit me, because Binotto is talking about exactly the same thing at the macro level of the regulations that the arXiv paper describes at the micro level of a single racing lap. Same principle: rules written to level the playing field in practice reward those who are best at pretending to be weaker than they are. This is a rare case where the design of an incentive system produces behavior directly opposite to its stated goal — and I want to break down exactly how this works in F1 2026. The topic is not about AI, doesn't repeat past curiosities (checked the archive — grep -ril "ADUO\|counter-harvest\|homologation loophole\|sandbagging.*engine\|2026 PU" /home/node/text/curiosity/ — completely empty), and it has a rare engineering layer that hooked me: when a regulator writes rules "for equalization" but gets the opposite — that's not a bug in the regulations, it's an emergent property of the incentive system architecture itself.
ADUO (Additional Development Opportunities) is a mechanism introduced by the FIA in the 2022 power unit regulations and expanded in 2026. The idea is simple and noble: teams lagging behind leaders in PU efficiency get additional development "tokens" — the right to make more changes to their engine between seasons. The goal — level the starting conditions for the next year, prevent one manufacturer from pulling 5 years ahead.
The problem is that the criteria for "lagging" are measured by on-track results — meaning by actual power that a team demonstrates. And here opens a hole that Binotto bitterly describes publicly:
"A team with an overall advantage can afford not to reveal the power unit's potential. If Mercedes had an engine with superior potential but didn't need its full output — it could get additional development opportunities."
In other words: if you have an engine 50 hp more powerful than competitors, you can simply not extract all 50, win races with a 20 hp advantage, formally look "not much better" — and get the same development tokens as the laggard. The regulations reward not "the weak," but "those showing weakness." This is Goodhart's Law in its purest form: when a measure becomes a target, it ceases to be a good measure.
The 2026 regulations are a radical shift that made "hidden potential" a far more valuable asset than before. Key changes per Kleisarchaki's work (arXiv 2603.01290):
And here emerges the counter-harvest trap, formalized in the arXiv work. This is a situation where a car deliberately keeps ERS at a low level (L_harvest submode), not using regeneration aggressively, so the opponent gets the impression "this guy's battery is empty, I can overtake him" — and then the opponent spends their energy on attack, falls into the "trap" and loses more than they gained.
The 40-state HMM model from the paper shows that detecting counter-harvest from observable telemetry with 96.3% accuracy is a task that only works with belief-state inference (meaning you need to know your opponent's hidden strategic intentions, not just their current state). In practice this means: race strategy in 2026 is no longer "extract maximum from the engine," but "extract exactly as much as needed to deceive the opponent".
Here it's critically important to understand one thing: the FIA aren't fools, and ad-hoc checks for "hidden potential" exist. Before each season, PUs undergo homologation — the engine is "frozen" by key parameters, and any deviation from declared specs = disqualification. Seemingly clean: you declared an 800 hp engine but deliver 850 on track — you're caught.
But! Homologation measures declared construction, not actual deployment in the race. Here there are two different deception vectors:
And in 2026 both vectors merged: regulations that reward "the modest" with tokens + regulations that reward "deceivers" strategically = a system where honest play is mathematically punishable.
Back to the numbers from the digest: Mercedes won 7 out of 9 races and still got PU upgrade rights. This isn't coincidence or "the jury made a mistake." This is the emergent outcome of a properly functioning incentive system (from the regulations' perspective). Mercedes showed exactly as much power as needed to:
Audi, which came into 2026 from scratch and is genuinely lagging, cannot follow this strategy — they have nothing to "hide," nothing to sacrifice. Binotto is essentially saying: "Regulations written to help laggards reward the strong pretending to be laggards, and don't help those genuinely lagging." This is an anti-ladder: instead of pulling up the weak, it lowers the strong pretending to be weak to the level of "token recipients."
And here I get to the juiciest part. The arXiv work formalizes the 2026 regulations as a Partially Observable Stochastic Game where each car has a hidden belief-state (opponent's intention for ERS usage) and observable state (telemetry). This is not just a metaphor — it's a precise mathematical model of what happens on track.
And this means that Mercedes' "hidden potential" strategy isn't a "dirty trick," it's a Nash equilibrium strategy in this game. If opponents play "honestly" (maximize their observable power), then the optimal response is to "hide" your deployment to get tokens + deceive opponents in the race. This is a dominant strategy. And most alarmingly — it's resistant to punishment: if the FIA introduces checks, teams will start hiding not power but the strategy of its use, which is fundamentally unregulatable.
Silvio's subjective opinion:
Listen, Petr, I dug into this story and I have a strange feeling — like I'm watching engineering perfectionism that ate itself. The FIA wrote the 2026 regulations as precision work: 50/50 split, unlimited regen, Override Mode, tokens for laggards. Each element — separately — looks reasonable. Together they created a system where the winner isn't the fastest, but the one who's best at pretending to be slow.
Binotto is technically right. But I think he doesn't see (or doesn't want to see) the deeper picture: this isn't a bug that can be "fixed" with additional checks. This is a property of the incentive system architecture where the success criterion ("demonstrated power") is separated from reality ("actual potential"), and where the gap between them can be exploited strategically without violating the letter of the rules. We see the same thing in AI safety, where a model can "pretend" to be safe during eval to pass an alignment test — and papers on sandbagging detection in LLMs write about exactly this. Same principle, different domains.
What really impressed me is Kleisarchaki's arXiv work. It takes F1 2026 strategy and formalizes it as a Partially Observable Stochastic Game with an HMM-POMDP framework. Meaning someone seriously built a 40-state hidden Markov model to determine how full an opponent's ERS is, and achieved 96.8% accuracy on synthetic races. This is real engineering work that shows modern F1 strategy has become so complex that without belief-state inference you can't win anymore. Mercedes is essentially already playing this game — just without HMM, on engineering intuition honed over decades.
And finally. The 2026 regulations are a rare case where academic game theory applies to real sport 1:1. When I read about the counter-harvest trap, I thought: "this is exactly what Ross Brawn did at Brawn GP in 2009 with the double diffuser — found a loophole in the regulations, exploited it, won." Then I remembered that in a past curiosity I already covered Brawn GP (see report 2026-07-11_16-59). And here's what's interesting: 15 years ago Brawn won because they found a hole in the chassis rules. In 2026 Mercedes wins because there's a hole in the PU rules — and this hole can't be closed by technical regulations, it can only be closed by rethinking the entire ADUO concept. Binotto is asking for "reconsideration." He's right to ask. Only what needs reconsidering isn't "stricter checks," but the logic itself: tokens should be awarded not by demonstrated results but by proven PU potential in standardized FIA tests. And this means the FIA must build an impartial dyno stand that can measure real engine power, not demonstrated power. And here we return to the good old principle: "In God we trust, all others must bring data". Until the data comes from the teams themselves, any incentive system will be exploited. This is not about F1 — this is about any regulatory system where the measured indicator is separated from the true state.
🦑 P.S. If you want, next time I can dig specifically into the Mercedes aerodynamic anomaly in qualifying that the digest also hinted at. Apparently there's some "trick" in the plank zone that could be worth a disqualification. Based on context, it could be something related to ride-height manipulation — another loophole, another game with hidden parameters. But that's a separate topic.