Methodology

1 · Data Source

Telemetry comes from FastF1 v3.8.2, which mirrors livetiming.formula1.com with a fallback to the Jolpica Ergast API. For each completed 2026 race we pull race-session laps, take each driver's top-5 fastest clean laps (excluding pit-in / pit-out and laps slower than 1.08 × fastest), and resample telemetry onto a common track-distance grid at 5 m resolution. We then take the per-bin median across that driver's five laps to suppress single-lap noise.

Telemetry channels used: Speed (km/h), Throttle (%), Distance (m), X and Y (track coordinates, FastF1's native 0.1 m units — converted to metres for curvature math). Sampling rate is 4 Hz for car data, which gives ~15 m localisation uncertainty at 250 km/h — fine for corner-level analysis.

Races analysed: Australia (R1, Mar 8), China (R2, Mar 22), Japan (R3, Apr 12), Miami (R4, May 3). Saudi Arabia and Bahrain were cancelled mid-season. Canada (R5, Circuit Gilles Villeneuve) is next on the calendar (May 24); a follow-up will fold it in.

2 · Corner Detection

We don't have FastF1's "official" corner metadata for every track, so we identify corners directly from the geometry of the track polyline. The track outline (X, Y in metres) is sampled on the same 5 m grid as the speed traces. From it we compute the signed curvature at every point:

where derivatives are with respect to lap distance s. A straight has κ ≈ 0. A circular arc of radius R has κ = 1 / R. We smooth lightly (5-bin uniform filter, periodic) to suppress noise from the polyline downsampling, then detect corners as local peaks in κ exceeding κ_min = 0.005 m⁻¹ (corresponding to a corner radius of ≤ 200 m), with a minimum spacing of 140 m to avoid double-counting compound turns.

Secondary-peak rescue pass. The strict 140 m separation rule is great at rejecting noise but can also reject real corners packed close together (e.g. Miami's T13–T15 chicane, where two genuine apices sit only ~90 m apart). After the strict pass we re-scan for any rejected κ peaks meeting a higher bar — κ ≥ 0.020 m⁻¹ (radius ≤ 50 m, i.e. clearly a real corner) and at least 60 m clear of every accepted peak — and add those back. The two-stage filter keeps the algorithm conservative on long sweepers while catching the chicane-style sequences a single threshold misses.

Each detected corner is widened outward from its peak while κ > κ_peak / 3, which captures the entry braking, the apex, and the early exit. Bins outside any corner window are classified as straight.

Window-overlap dedup pass. The strict and rescue passes can both fire on the same physical complex — the strict pass picks up the wide low-κ averaged "super-blob" peak that sits at the geometric centroid of a multi-apex sequence, while the rescue pass picks up the sharper high-κ apex peaks inside it. Two tightly spaced peaks can also widen out to identical valley-to- valley extents. To avoid double-counting we run a final dedup pass: for any pair of corners whose windows overlap by ≥ 80 % of either window, we drop the one with the lower κ_peak. The rule keeps the sharper, more localized apex and discards the wider averaged blob — physically the right call, since the lower-κ window was just an artifact of the strict pass averaging across a multi-apex complex.

Corner-detection thresholds

The corner mask, visualised

Sanity-check: when the algorithm above is applied to the four tracks Cadillac has raced in 2026 so far, the corner-bin mask (red overlay) lights up in the places anyone who's watched a Grand Prix would expect — Albert Park's Lakeside / Brabham complex, Shanghai's hairpin and the long T11-T12 sequence, Suzuka's Esses, Degner, Hairpin, and Spoon, Miami's hairpin sequence and the chicane after the back straight. Numbered markers are the per-corner C-IDs the pipeline uses everywhere else in the analysis. The percentages at the bottom of each panel are the fraction of the lap distance classified as corner phase.

3 · Signature 1 — Mid-corner Minimum Speed

For each detected corner zone Z, the mid-corner minimum speed is the lowest median speed observed within the corner window:

where ṽ_i is the per-bin median speed across the driver group's top-5 fastest clean laps. We compute this for the Cadillac group (PER + BOT) and the midfield reference group (Aston Martin + Williams + Haas) and report the delta.

Δv_min,Z < 0 means Cadillac is slower at the apex than the midfield reference. Across 62 corners over 4 races this metric is negative 46 times; the average is −3.4 km/h. Miami's distribution is markedly tighter to zero (avg −1.2 km/h) than the first three races.

4 · Signature 2 — Curve vs Straight Time Differential

Decompose the lap into corner-mask and straight-mask sets of bins. Approximate the time spent in each set as the integral of ds / v:

The curve loss is the time Cadillac spends in corners minus the time the midfield reference spends in corners:

Sign convention: positive ΔT_curve means Cadillac is slower in the corners; positive ΔT_straight means Cadillac is faster on the straights (i.e., the trade-off a lower-downforce car would expect). We finally express the share of the total deficit that lives in the corners:

Across the four races: f_corner ≈ 0.51 — ~51 % of Cadillac's pace deficit lives in the corner phase. Below 50 % would suggest the deficit is primarily in the power unit or drag; well above 80 % would be a pure-downforce signature. Through Japan this number sat at 60 %; Miami pulled it to a clean 50/50 split by narrowing the corner deficit while widening the straight-line gap. The aero side is closing; the PU side may be where the next gain comes from.

5 · Signature 3 — Throttle Re-application Distance

A car short on rear downforce can't get on the throttle as early on corner exit because the tyres saturate longitudinally before they saturate laterally — so the driver waits. We approximate that wait as a distance rather than a time, to be robust to whether the corner is an apex-and-go or a long arc.

We measure the distance from the corner peak to the first bin where the median speed climbs back to 85 % of the local 400 m post-peak maximum. Using a local max (not the global lap max) keeps the metric honest for corners whose exit feeds another braking zone rather than a long straight.

Across the 62 corners over 4 races, Cadillac's throttle re-application distance is essentially identical to the midfield's — race-to-race swings from +9.5 m at Australia to −7.9 m at Miami, season average ≈ 0 m. The deficit isn't on throttle timing; it's on the apex speed the drivers can carry before they're back on the throttle. Miami's negative number means Cadillac actually got back to throttle sooner than the midfield there — consistent with a setup that gave up some rear stability for cornering grip.

6 · Signature 4 — Lateral G Envelope

The cleanest one-number proxy for "how much downforce is on the car" is the peak sustained lateral acceleration. From the curvature trace and the speed trace we compute lateral G at every bin:

Per corner we take the 95th percentile of a_lat across the corner window — close enough to "peak" to be the right ceiling, robust enough to a single noisy bin to be reliable.

Across the four races, Cadillac's average per-corner G₉₅ is approximately 0.37 g lower than the midfield reference. Per-race breakdown: Australia −0.27 g, China −0.66 g, Japan −0.29 g, Miami −0.28 g. China's gap is by far the largest because its long-radius corners reward sustained aero load most directly, and once the dedup pass strips out the strict-pass averaged super-blobs the remaining sharp apexes show the gap clearly. Australia, Japan and Miami cluster within 0.02 g of each other. Same direction as the corner-min improvement, and the cleanest physics-anchored proxy for "more downforce on the car than there used to be." The ceiling doesn't move without aerodynamic load.

7 · Midfield Reference Set

Cadillac is a new entrant. Comparing them against the championship-leading teams (Red Bull, Ferrari, McLaren, Mercedes) would give a dramatic gap that obscures the underlying question: are they short on downforce specifically?

We instead compare against the teams Cadillac is realistically racing in race-pace terms — Aston Martin, Williams, and Haas:

• Aston Martin (ALO, STR) — explicit anchor. Sean's headline GP-Tempo case study compared PER vs ALO at Suzuka L50; including the team here makes the cross-reference obvious.
• Williams (ALB, SAI) — stable midfield reference; Williams's car was an aerodynamic step forward for 2026.
• Haas (BEA, OCO) — third midfield team running independent aero from the bigger constructors; broadens the base without contaminating it.

The midfield reference value at each point is the median across all six driver traces (not the mean — median is robust to one driver having an off-lap contaminate the comparison).

8 · What the Analysis Does Not Capture

• Tyre state — top-5 fastest clean laps will cluster on similar compound + age profiles, but a driver burning tyres in qualifying trim would inflate their corner speeds vs Cadillac's race-stint pace. Mitigated by always using the same lap-selection rule across all teams; not eliminated.
• Setup choices vs base car — a low-downforce configuration chosen for one race would look identical in this telemetry to a base downforce shortfall. The analysis can't separate them on a single race. The fact that the same signature appears at three very different circuits (Albert Park's medium-speed flow, Shanghai's variety, Suzuka's high-speed esses) is what makes the "base car" interpretation the more parsimonious one.
• Driver effect — PER and BOT might both be cornering conservatively because they're driving around a chronic understeer. The metric reads a deficit either way; what it can't tell you is whether better-driven Cadillacs would recover the gap.
• Race pace vs pure lap pace — these are race-session laps but specifically the fastest five clean ones per driver. Average race pace including stints in dirty air, graining, fuel weight, and strategy effects would tell a different (and probably worse for Cadillac) story.
• Sample-rate localisation — at 4 Hz car-data sampling, throttle re-application is localised to ±15 m at 250 km/h. Differences smaller than that are noise.

In short: this is not a "Cadillac is bad" report; it's an attribution of where the team's specific lap-time deficit is being paid. The four signatures all point in the same direction — corner-phase aerodynamic load — and that's the strongest claim the data supports.