# The setup sheet, baseline philosophy and back-to-back testing

> A setup is only as good as its documentation. Keep a complete setup sheet and run log, work from a known baseline, evaluate changes back-to-back with controlled conditions, and check whether a lap-time difference is statistically real before believing it.

Area: Setup & Driver Feedback · Starting level: L2 Practitioner · Updated 2026-11-16
Source page: https://race-engineer-academy.vercel.app/lesson/setup-sheet-baseline-testing

## L1 · Foundation

**The setup sheet** is the single source of truth for the car's configuration. It records every adjustable: tyre pressures (cold set, hot achieved), ride heights, corner weights, camber, toe, caster, springs, ARBs, packers/bump rubbers, damper clicks, wing/flap positions, brake bias, diff settings, electronics maps (TC, ABS, engine), gear ratios, fuel load and ballast. Alongside it, the **run log** records conditions (track/air temp, wind, time), tyre set and age, fuel, the driver's comments and lap times.

**Baseline philosophy.** Every event starts from a **baseline**: a known, validated setup the team trusts (from the last visit, simulation, or the manufacturer's recommendation). Changes are evaluated *relative* to it. If the weekend goes wrong, you can always go back. Never start "from scratch".

**Back-to-back testing.** To know whether a change helped, compare it with the baseline under conditions as similar as possible: same driver, same tyre state, similar fuel, close together in time. The cleanest method is **A–B–A**: baseline, change, baseline again. If the second A matches the first, conditions were stable and the B result is believable.

**Is the difference real?** Lap times scatter by 0.1–0.3 s even for a consistent pro driver. A 0.1 s improvement on one lap means nothing. You need several laps per configuration and you must compare the difference with the scatter.

## L2 · Practitioner

**Practical documentation**
- One sheet per run, version-numbered (e.g. "FP2-R3, baseline + front ARB +1"). Mark changes since the previous run clearly.
- After every run: hot pressures, tyre temps, ride-height check if a kerb strike or off occurred, fuel used, driver comment verbatim, your interpretation, and the next action.
- Record *measured* values, not just intended ones: a click count is meaningless if a damper was rebuilt; corner weights after a spring change can move by 5–10 kg.
- Keep a **change log of the car hardware**: parts mileage, damper rebuilds, gearbox ratios – unexplained balance changes are often hardware (a broken ARB blade shows up as a non-linear roll ratio, Segers §9.6).

**Run planning for back-to-back**
- Control tyres: evaluate both configurations on the same set and similar age; avoid comparing a new-tyre lap to a 10-lap-old tyre lap.
- Control fuel: either equal fuel or correct with a measured fuel sensitivity (order of 0.02–0.04 s per kg for a ~1:40–2:00 lap, car-dependent – measure it).
- Control track evolution: ABA or ABBA ordering; note wind and track temperature.
- Use **segment times and metrics** (corner minimum speed, sector times, average understeer angle in a phase) rather than only lap time – they have lower noise and point at *where* the gain is.

**Significance quickly.** With lap-time standard deviation $\sigma$ and $n$ laps per configuration, the standard error of the difference of means is $\sigma\sqrt{2/n}$. A rule of thumb is to believe a difference larger than about twice that.
Example: $\sigma = 0.15$ s, $n = 5$ → SE = 0.095 s; a 0.08 s difference is **not** resolved. To resolve 0.1 s at ~95% confidence you need roughly $n = 2(2\sigma/\Delta)^2 = 18$ laps per configuration. That's why many small changes are better judged with metrics, the driver's feel, or simulation – and why large steps are tested first.

## L3 · Advanced

**Statistical treatment of lap-time comparisons** (cf. Segers Ch. 17, metrics and run charts).

1. **Clean the samples**: remove out/in laps, traffic, yellow flags, track-limit offs. Use objective filters (e.g. laps within ~101–101.5% of the stint best, or within 2–3σ of the stint median; no lift > x% on full-throttle sections; no track-limit or yellow-flag laps) defined *before* looking at the results.
2. **Correct for known drifts**: fuel burn (s/kg × kg per lap), tyre degradation (fit a linear trend within each stint), track evolution (from ABA: interpolate A between its two runs).
3. **Test**: Welch's t-test on corrected lap times, $t = \Delta / \sqrt{s_A^2/n_A + s_B^2/n_B}$. With small $n$ use the t-distribution (for 5+5 laps the 95% critical value is ~2.3, not 1.96).
4. **Power**: to detect effect $\Delta$ with significance $\alpha = 0.05$ and power 80%, $n \approx 2\left((z_{1-\alpha/2} + z_{1-\beta})\,\sigma/\Delta\right)^2 = 2(2.8\,\sigma/\Delta)^2$. For $\sigma = 0.15$ s and $\Delta = 0.1$ s: $n \approx 35$ laps each – impossible in a normal session. Conclusion: lap time alone cannot resolve small setup effects.

**Lower-noise metrics.** Sector or corner-segment times have smaller $\sigma$ (noise doesn't add up across the lap). Physical metrics often have better signal/noise: minimum speed in a corner, time to full throttle, average understeer angle gated by phase, roll gradient or roll ratio (Segers §9.3–9.5: roll gradient per lap is a direct check that the bar change actually changed the car), grip factor. These can confirm the *mechanism* of a change even when lap time is inconclusive.

**Baseline management.** Re-run the baseline periodically (e.g. start of each test day) to detect drift from hardware (damper fade, worn bushes, tyre batch). Keep a **reference lap** and a **reference metric set** for the baseline per track. When a change is adopted, the new baseline must be documented as a full sheet, not as "old + changes" spread across several pages – reset errors are a frequent source of wasted sessions.

**Confounding by the driver.** The driver knows what has been changed and adapts. Blind tests (the driver not told which configuration) reduce expectation bias for subjective ratings. Two-driver crews: evaluate with both, because a change can help one and hurt the other (Segers Fig. 7.28).

## L4 · Expert

**Designing tests like an experimentalist.** RCVD's testing chapter frames development as controlled experimentation; modern teams extend it with simulation-in-the-loop.

- **Blocking and randomisation.** Track evolution is a monotonic nuisance; ABBA or ABAB with interleaved short runs cancels linear drift and reduces the effect of driver adaptation (both appear in both orders). Tyre state is a second nuisance: run each configuration on the same set in alternating fashion, or block by set.
- **Factorial designs.** When two parameters interact strongly (e.g. rear ARB × diff power ramp, rake × rear wing), a 2×2 factorial (4 runs, or 5 with a repeated centre point) estimates both main effects and the interaction with the same laps that two one-factor tests would use. Fractional factorials scale this to 3–5 factors on a simulator or a long test day.
- **Bayesian updating of the baseline.** Treat the baseline's expected lap time and balance metrics as priors from simulation and previous events, and update with each run. This formalises "how surprised should I be?", e.g. a result that contradicts sim by 3σ is more likely a measurement or reset error than a discovery – check the car first.
- **Correlation with simulation.** Use each run to correlate the vehicle model: compare logged roll gradient, ride heights, understeer angle vs $a_y$, and speed traces against the sim. Systematic deltas point to model error (tyre model, aero map, compliance). Setup decisions are then made on a correlated model, with track running used to confirm, not to search.
- **Decision under uncertainty.** If the evidence for a change is weak but the expected effect is positive and risk is low, adopt it provisionally and keep monitoring – waiting for 95% significance in a race weekend means never changing anything. Conversely, a change that improves one-lap pace but increases variance or deg may be rejected for the race.

## Key formulas

**Standard error of a lap-time difference**

$$SE_{\Delta} = \sqrt{\frac{s_A^2}{n_A} + \frac{s_B^2}{n_B}}$$

- s: lap-time standard deviation per configuration (s)
- n: number of clean laps per configuration

**Laps needed per configuration (α = 0.05, power 80%)**

$$n \approx 2\left(\frac{2.8\,\sigma}{\Delta}\right)^2$$

- \sigma: lap-time scatter (s)
- \Delta: difference to detect (s)

## Trackside takeaways

- Document everything, every run: measured values, conditions, driver words, your interpretation, next action.
- Always start from a validated baseline; re-run it to detect drift.
- Use A–B–A (or ABBA) to cancel track evolution and driver adaptation.
- A lap-time difference is real only if it clearly exceeds ~2× its standard error.
- Prefer lower-noise metrics (segment times, min speed, gated understeer angle, roll gradient) to judge small changes.
- Correct for fuel, tyre degradation and track evolution before comparing.

## By car category

### GT3

Multi-driver line-ups: log which driver drove each run; compare like with like. Data sharing across sister cars helps build baselines.

### One-make Cup

Same car, same tyre for everyone: small, well-documented setup gains and consistent tyre preparation are the differentiators.

### FIA F3

Limited testing and short sessions: baseline quality from simulation and previous events dominates; plan very few, big-step changes per session.

### LMP2

Long test days allow ABA and factorial work on the ORECA; keep strict reset procedures (ride heights and corner weights on a flat patch) between runs.

## References

- Milliken & Milliken, Race Car Vehicle Dynamics (SAE 1995): Ch. 11 Testing and Development
- Milliken & Milliken, Race Car Vehicle Dynamics (SAE 1995): Ch. 12 Chassis Set-up
- Segers, Analysis Techniques for Racecar Data Acquisition, 2nd ed. (SAE 2014): §17.1–17.3 Metrics and run charts
- Segers, Analysis Techniques for Racecar Data Acquisition, 2nd ed. (SAE 2014): §9.6 Suspension Troubleshooting
