Skip to main content
Every model in the project runs on “usable” laps, and the rest are deliberately thrown away. The data quality audit makes that visible: for each race in a season, how many laps passed the usability gate, and exactly which lap types were excluded and why. It is the data-engineer’s view of the pipeline the honest denominator behind every other chart.

​
What it shows

Each race is broken into a usable share and exclusion buckets:
  • Usable %: the fraction of laps flagged usable_for_modelling = true. A lap is usable when it is not a safety car / VSC period, not an out-lap or in-lap, not rain-affected, and carries no major anomaly flag.
  • Neutralised: laps under safety car or VSC (correction_class = 'neutralisation'); pace is unrepresentative, so they are excluded from degradation and skill models.
  • Rain: laps where wet weather was detected; degradation behaviour is qualitatively different and excluded from dry-compound models.
  • Out / In laps: the lap after a pit stop (cold tyres) and the lap into the pits (lift-and-coast), both excluded from degradation slopes.
  • Anomalies: laps with anomaly_class != 'normal' mechanical events, mistakes, or clean-cliff detections that carry special handling.
The buckets are exhaustive of the exclusions, so usable % plus the excluded categories account for every lap in the race.

​
How to use it

  1. Select a season from the filter bar; the per-race breakdown loads.
  2. Scan usable % to spot races with thin modelling data (wet events, safety-car-heavy races).
  3. Read the exclusion buckets to see why a race is thin rain versus neutralisation tells very different stories.
  4. Use it as a caveat layer a feature’s confidence at a given race should be read against that race’s usable %.

​
Reading the signal correctly

A low usable % is not a data error it is the pipeline correctly discarding unrepresentative laps. A wet, safety-car-interrupted race should have few usable laps; the audit confirms the gate is working, not that data is missing.
Races with a very low usable % feed every downstream model less data, so their estimates elsewhere in the app carry wider uncertainty. Treat this page as the provenance check behind the rest of the suite.

​
Data source

The chart queries int_lap_anomaly_flags via DuckDB-Wasm in the browser, aggregated per race across a full season. This is a full-season view across 2018–2024 data. For the full model definition, see the int_lap_anomaly_flags reference.