The synthetic archive and the reading list
Where the course's data comes from, and the real sources its claims rest on
The archive#
Every race record, lap time, forecast and result in this course is generated by the Faculty. None of it is recorded from a real season, and no real timing data is used anywhere.
A measured archive has no true value to hand back, so an estimate from one can be admired but not checked. A generated archive has the true values behind it, which is how this course can tell you that you are wrong.
What it hands back changes across the weeks.
- Given. In the first act the true value arrives with the data. Week 2’s Lab is handed a coefficient no single stint could have told it, so its estimate can be put beside the thing it was estimating.
- Withheld, then revealed. Through the middle weeks you commit first and the truth is released afterwards. That order is what lets a week be built to overturn the one before it: week 5’s scenario is constructed so the stop lap that wins on time and the one that wins on position differ, and a student’s own week-4 model is what shows it.
- Released in stages. In week 11, supplied race states appear during the two timed runs, but the model’s ranking and therefore call correctness stay hidden until both runs finish. In week 12, a condition-redacted archive opens after validation for the aggregate check; condition labels and the remaining ground truth open only after that checkpoint, for the stratified audit. The supplied-model domain is assessed in the lab. Its blank audit plan, not that domain, is what students apply to their own instrument.
The reading list#
Claims about the real world are different, and this is where their support is visible. The sources below carry them, each only in part.
FIA 2026 Formula 1 Regulations, Section B — Sporting, Issue 08. This pinned edition supports the tyre allocation, what counts as using a set, returns, the tyre change permitted during a suspension, safety-car procedure and race classification by finishing position. It does not support circuit-specific tyre selections, the course’s traffic values, a safety-car arrival probability or a universal pit-loss reduction.
FIA 2026 Formula 1 General Regulatory Provisions, Section A, Issue 03. This separate pinned edition supports the two-car entry, the Constructors’ Championship calculation from both cars’ results, and the race points schedule used for a race reaching at least 75% distance. It does not make the course’s expected-points calculation an FIA recommendation and does not support Sprint, shortened-race, tie-break or penalty claims here.
Aguad and Thraves (2024). They support modelling pit strategy against a rival who responds, the feedback Stackelberg and dynamic-programming framing, and the distinction between deterministic and stochastic formulations. They supply no classroom number, and the bounded table of rival responses used in teaching is not their algorithm.
Heilmeier et al. (2020). They support sampling probabilistic effects inside a lap-wise race simulation, modelling safety-car phases, and producing a distribution of final ranks across repeated runs. They do not supply the course’s hazard, pit costs, sampler, outcome map or calibration claim.
ECMWF, “Quantifying forecast uncertainty”. It supports describing an ensemble as multiple forecast realisations for the same time and place, created by varying initial conditions and model representations. It supplies no Formula 1 rain path, tyre effect or course weight.
Pearce, Poloczek and Branke (2022). They support the shared-scenario operation behind common random numbers: reuse a seed so alternatives face the same generated input. They do not establish that pairing always reduces variance, removes sampling uncertainty or calibrates a race probability.
Guo et al. (2017). They support fixed-bin counts, observed frequencies and the reliability diagram’s identity reference. They do not supply the course’s event, bins, races or dry/wet behaviour.
Hébert-Johnson et al. (2018). They motivate checking calibration within an identifiable set instead of letting aggregate agreement settle every conditional claim. This course borrows that diagnostic lesson, not their fairness setting, algorithm or guarantees.
What no source carries. Every classroom race record, forecast weight, position, time and result is synthetic. The archive’s dry and wet behaviour is a property of its generator, attributed to nothing else. Week 12 makes the same point: agreement inside an archive is not a guarantee about a race that has not happened.