Weigh driving against setup, with the evidence.

Racing's AI reads the synthetic Jerez comparison and writes competing driving and setup hypotheses, each tied to the evidence it cites and to what is missing. You correct the context, request a revised analysis and leave with one editable test plan. It advises; engineering judgment stays with you.

Engineer’s debrief

AI-generated, advisory

Evidence, competing explanations, and a test to separate them.

Driving explanation

Run B lifts earlier and returns to throttle later through the corner window, which may explain the slower exit.

  • B02 corner window
  • A02 corner window

Setup explanation

The front roll bar moved from position 3 to 4 in run B, which may add entry understeer. These traces alone cannot establish the cause.

  • Run B context
  • Run A context

What remains uncertain

Hot tyre pressures were not measured in either run. Adding them as context would narrow the setup explanation.

Missing evidence

  • Hot tyre pressures, run A
  • Hot tyre pressures, run B

Proposed next test

Repeat before changing the car

Hold the setup fixed. Repeat run A's lift and throttle cue on clean laps. Compare the corner window and run medians.

Example wording from the synthetic scenario, not a live AI result.

Run B against run A

Jerez · Fictional GT4 #27 · Speed (km/h) · Lap distance (m), sector 2

Synthetic scenario
Median lap delta
+0.780s
A02 / B02 · sector 2
+0.580s
Synthetic Jerez sector 2 speed comparisonRun A lap A02 is a dashed grey reference line and run B lap B02 is a solid red line. Speed in kilometres per hour against lap distance in metres. Both slow for the corner between 2100 and 2800 metres, where run B carries less speed on exit.Corner window200150100501400 m1800 m2200 m2600 m3000 m

Reading the chart

Run A, A02
Reference lap from run A. Front roll bar at position 3.
Run B, B02
Compared lap from run B. Front roll bar moved to position 4.
Corner window
2100 to 2800 m. The window the evidence cites.
View speed data as a table
Speed every 100 m through the corner window, read from the chart samples and rounded to the nearest km/h.
Distance (m)Run A, A02 (km/h)Run B, B02 (km/h)
2000166166
2100166166
2200166163
2300131127
24007573
25007573
2600131127
2700167163
2800167167
Figure 2. The traces the debrief cites, from the same synthetic scenario.Representative laps stay separate from run medians.

What the AI race engineer does today

Everything here ships today on the synthetic scenario. The model runs on the OpenAI API and is called only when you request an analysis.

  1. Reads the evidence you selected: run medians, the laps you kept, the A02 and B02 corner windows and each run's reported context.
  2. Writes driving and setup hypotheses, each with the evidence it cites, what remains uncertain and which measurements are missing.
  3. Proposes one next test that can tell the explanations apart.
  4. Takes your corrections. Add what the data cannot show, then request a revised analysis. Earlier results keep the revision they were built on.
  5. Turns the proposal into one editable test plan you save: the change, what to hold constant, the measurement, the success criterion and the stop condition.

What it does not do

It interprets measured differences; it does not prove a cause. Every result carries the label AI-generated, advisory.

  • It does not fill a missing measurement with a guess or a zero.
  • It does not claim a lap-time gain for any change.
  • It does not choose the test for you. The plan is yours to edit before you save it.

What is sent, and when

Only when you request an analysis, the selected evidence and your saved context notes go to the OpenAI API with storage disabled. The provider may keep requests for up to 30 days for abuse monitoring. Data ownership shows the full flow.

The method behind the debrief

The debrief follows the order an engineer would. Driving or setup: separating causes explains how to read the traces. Designing one discriminating test explains how the next test is chosen.

Design-partner roadmap, 2027 season

Planned with design-partner teams for the 2027 season. Not available today.

  • A per-team isolated model trained only on that team's data. No cross-team training and no shared model.
  • MoTeC and AiM CSV import, so the debrief reads your own logged sessions.
  • Team data stays the team's, with export and deletion when a team leaves.

Read the debrief on your own screen.

Private early access. Request an analysis of the synthetic Jerez comparison and inspect every citation.

Request access