Computed vs. generated: how PitDelta decides what you practise

Most sim-racing analysis tools now describe themselves the same way: upload your telemetry, AI finds your mistakes. If you have wondered what is actually deciding what you should practise — a set of rules someone can be held to, or a language model’s best guess — this article is the straight answer for PitDelta. It only describes what PitDelta does; other tools can speak for themselves.

What "computed" means here

When you upload an iRacing .ibt file, everything that matters in the debrief is produced by encoded telemetry rules: which corners become zones, how much time each one costs, what mechanism is behind the loss, how a faster driver’s session decomposes against yours, and which single job is worth your next stint. A language model does not read your telemetry, does not pick a zone, and does not choose the assignment.

That has a practical consequence you can test: run the same telemetry with the same comparison context on the same engine build, and you get the same analysis. Not "similar advice, phrased differently" — the same zones, the same numbers, the same assignment. Reproducible does not automatically mean correct; it means the analysis can be checked, repeated and argued with, which is what makes the rest of this article possible to verify.

The one AI-written line, and its fence

There is exactly one AI-generated element in a PitDelta report today: the engineer’s one-line remark. It is written from the already-computed summary — the same track name, lap counts and per-corner findings printed on the card. Raw telemetry is never sent to the language model. The remark cannot select a zone, change a number, or pick the practice assignment; if the line fails validation or the model is unavailable, a fixed fallback line is used and the analysis is untouched. And PitDelta never uses your telemetry to train AI models.

Why the report sometimes refuses

Not every difference becomes advice. PitDelta looks for what repeats across your clean laps before it tells you what to change, and when the evidence is weak, contradictory or poorly matched, it says so instead of inventing coaching:

That refusal is engine behaviour, not an editorial choice — the same gates decide what a crew room will and will not compare, which is why a wrong-car or wrong-configuration file is excluded with a note rather than quietly skewing the board.

If you have ever taken confident-sounding advice into a stint and made your driving worse, you already know why this matters. A tool that always has something to say is easy to build. Withholding is the harder promise.

Learning versus repeating

The clearest example of computed analysis doing something a lap-time delta cannot: with a faster driver’s file, PitDelta separates the part of the gap you still need to learn from pace you have already shown but do not repeat. One calls for a technique change; the other calls for a repeatability drill. Those are different practice problems, and treating the whole gap as one number sends you to the wrong one. The report defines the split in its own words:

Check it yourself

Every report states its own basis in the footer — the number of clean laps it was computed from, the engine version and the build, in the form "Computed from 14 clean laps · Engine 7 · build f8369b2". That line is the handle for the reproducibility claim above: same telemetry, same comparison context, same build printed in the footer, same analysis.

In short

The practice assignment is computed from telemetry by encoded rules and withheld when the evidence does not support it. One line of the report is AI-written, from the computed summary only, behind validation and a fallback. Everything in this article is a description of PitDelta’s own shipped behaviour — upload a session and check it against the report you get back.