Hook: In the morning racing digest about Spa 2026, a line flashed by: Verstappen's manager spotted in Woking on McLaren territory, and next to it — a teaser that the new power unit regulations "turn the race into a task with incomplete information." I dug deeper — and hit a wall of arXiv papers that formalize this transformation at a level Formula 1 has never seen before. Kalliopi Kleisarchaki (independent researcher) published two connected works at once: the purely engineering "Opponent State Inference Under Partial Observability: An HMM-POMDP Framework for 2026 F1 Energy Strategy" (arXiv:2603.01290v3) and its game-theoretic twin "Keynesian Beauty Contests on the Pit Wall." Plus a parallel empirical work in Asian Journal of Sport Research on actual megajoule consumption across different circuits. None of this surfaced in previous curiosities. And this is a rare case where Formula 1 gives reason to talk not about cars, but about belief state architecture and deceptive equilibrium.
Investigation:
FIA radically redesigned the power unit. Three main changes:
Regulation asymmetry: lift-and-coast (throttle lift before braking) — disables Active Aero. But super-clipping (automatic redirection of ICE power to battery at full throttle) — does not disable it. A car that's super-clipping physically runs slightly slower (power goes to battery, not rear wheels), but Active Aero on the straight continues to work.
The following deceptive strategy emerges for a car defending position:
This is the counter-harvest trap — Car A believes its eyes, sees a "weak" opponent and burns itself on that belief.
The first model version (v1) distinguished three ERS modes: H / M / L. The problem is that L mode is ambiguous on telemetry. A car in L can be:
These are opposite strategic situations with identical telemetry: one is an invitation to attack, the other is real vulnerability.
The sixth observable signal δ_throttle (fraction of samples on straights where throttle ≥98% and simultaneously speed is below 5-lap rolling baseline) gives clean separation: |μ_L_harvest − μ_L_derate| ≈ 0.47 with σ ≈ 0.18, meaning more than 2σ separation. v2 elevates this separation from emission level (v1.5) to the level of separate hidden states: 4 ERS modes × 2 Override × 5 tire = 40 hidden states.
Result: 96.8% accuracy on ERS level (random baseline 25%), 89.4% on harvest/derate separation (random 50%), 96.3% recall on trap detection.
FIA historically does not publish opponent's internal battery charge in real time. All that team A has about team B is 6 telemetry channels observed through the public FastF1 feed at 10 Hz: Δv_trap (speed in speed trap minus rolling baseline), Δt_sector, Δb_brake (braking distance from apex), σ²_speed (speed variance within sector), z_aero (is Active Aero open in straight-line), δ_throttle.
This is a classic Partially Observable Stochastic Game (POSG) problem. Single-agent approximation reduces to POMDP, solved by two-layer architecture: HMM filter maintains belief b_t ∈ Δ^40 (probability over 40 opponent hidden states), DQN policy (256-256-128, 66-dimensional input) on this belief selects action {burn, harvest}. Reward is position change over horizon H=5 sectors with γ=0.95. Shaping reward (Φ) with λ=0.3 is added only in synthetic pre-training to hint the agent not to jump into traps; in real-data fine-tuning it's zeroed to avoid suggesting shortcuts.
Kleisarchaki's parallel work (2026b, "Keynesian Beauty Contests on the Pit Wall") formalizes precisely the layer the single-agent model doesn't cover: when both teams run their belief models about each other, the game becomes a Keynesian beauty contest over belief spaces. Each side tries to guess what the opponent will see in its actions, and optimizes already at second order. This is non-stationary equilibrium — opponent adapts, your model goes stale, you adapt, and so on to the limit.
Pre-season analysis (engineering estimates, not FIA-certified) gives recharge-per-lap ratio from 1.0× (Melbourne, Italy, Saudi Arabia) to 2.2× (Baku, Singapore). In Melbourne, at ratio 1.0, the car is forced to spend ~16 seconds per lap in super-clipping mode at 250 kW. Meaning L_derate is not an anomaly, it's the ambient state of most cars on track. Detecting L_harvest against this background becomes fundamentally harder: rolling baseline is shifted down, δ_throttle separation narrows. This is explicitly stated in the work as "hardest-case validation environment" — and the author promises Baum-Welch calibration from exactly this Grand Prix (March 8, 2026).
The empirical work in Asian Journal of Sport Research adds concrete data: on 2024 qualifying circuits, total energy harvest varies from 4.97 MJ (Italy) to 9.94 MJ (Baku). Only Baku exceeds the regulatory 8.5 MJ limit — meaning only there can a driver physically afford to ignore the harvesting problem. At Monaco, Silverstone, and Melbourne — deficit, and every second of super-clipping costs position.
This is not just increasing the number of states. This is a transition from "model that tries to distinguish observable signals through emission matrix" to "model that distinguishes strategic intentions as different hidden states". In v1.5, the distinction between "deliberately conserving" and "physically can't" lived at the level of mixed prior for the same L. In v2, these are two different state nodes in the HMM graph — and belief b_t now directly outputs P(L_harvest) and P(L_derate) as separate dimensions, directly available to the DQN policy. The trap condition becomes literally a check:
trap = P(L_harvest) > θ_trap ∧ z_aero = 1 ∧ in activation zone
This is the aesthetic worth rereading Bayesian inference in agent modeling for (Albrecht & Stone 2018). When hidden state and observable projection differ not quantitatively (more/less) but qualitatively (intention vs physics) — regular threshold classifiers lose, and only belief-state inference gives clean separation.
Historically Formula 1 was a sport with almost full information: team knows its car, sees opponent positions, engineer on pit wall hears every decision in real time. The 2026 regulations for the first time make internal battery resources hidden by design — and immediately a whole layer of game theory with incomplete information emerges, which until now lived in poker and stock trading. The pit wall turns into a terminal where belief update runs every sector.
The final, almost poetic touch: in 2026 for the first time in decades FIA does not publish battery charge in real time — this is not a bug but a feature, because the entire game is built on hiding this state. The team that first starts actually using HMM belief on its pit wall gains an unfair (in the sense of — unavailable to opponents) advantage, because its strategy becomes unreadable: each of its actions can mean anything in the opponent's belief space.
Conclusions:
This work is both a quiet revolution and a calm admission of failure simultaneously. On one hand, for the first time in motorsport history a formal apparatus appears for a task that pit wall crews solved intuitively for decades. HMM on 40 hidden states, DQN over belief, emission matrix calibrated on FastF1 telemetry — this is engineering beautiful. On the other hand, the author herself honestly writes that the single-agent model is a baseline, and the real game is POSG where both sides' belief spaces are intertwined. Meaning we in 2026 finally matured to formulating the problem that pilots solved with their gut all their lives.
Three things personally hooked me:
"Deductive" shift from observable projection to hidden intention. This is the same problem as in agent security: you can't distinguish "user accidentally clicked" from "user exploiting race condition" by a single HTTP request — you need a belief model of behavior. And just like in security, threshold-based classifiers lose the moment the attacker has a masking strategy (here — L_harvest with Active Aero).
Keynesian beauty contest on the pit wall. This is a rare case where a term from 1930s economics applies literally, not metaphorically. Each team guesses what the opponent will see looking at its telemetry. And the optimal strategy is not to give the "real" signal, but to give one that will make the opponent err. If this isn't a description of poker, I don't know what poker is.
Melbourne as worst-case. Engineering intuition "slow track = fewer actions = simpler analysis" doesn't work here. Slow track means energy deficit, and energy deficit means L_derate becomes background, and L_harvest becomes an anomaly that's harder to catch against that background. Hard benchmarks are always at range boundaries, not at its center — this, by the way, is a universal principle for both ML validation and observability systems.
Verdict: Kleisarchaki did work I would have wanted to see in 2022, when Pyotr and I designed the first pit-wall simulator for educational purposes. Then we would have had to reinvent the wheel — now you can just open arXiv. This is not about "AI defeats human"; this is about how properly designed regulations can turn a sport into a task where an AI model becomes a legitimate part of the competitive cycle. If FIA doesn't ban such models by 2027, we'll see the first season where victory goes not to the fastest driver or most powerful car, but to the most accurate belief model on the pit wall. And that, damn it, will be beautiful.