Data collection coverage is how well collected data represent the situations and conditions a system must handle. A narrow, predictable operating domain can approach complete coverage; a broad, uncontrolled domain creates missing cases and lower capability. section 9.10 section 9.9
Coverage follows an S-curve: capability improves as coverage grows, then delivers diminishing returns. If the domain is too broad, capability may level off well below 100%, forcing a narrower objective or a smaller set of viable capabilities. section 9.10
Coverage gaps become especially dangerous when systems encounter conditions missing from the data:
- A warehouse forklift operates in a controlled setting with predictable lighting, uniforms, and routes, making high coverage more achievable. section 9.10
- An urban self-driving vehicle faces uncontrolled variables and unusual edge cases, such as a pedestrian wearing a chicken costume that never appeared in training data. section 9.10
- The Challenger O-ring tests contained no performance data below 40 degrees Fahrenheit, leaving a critical hole before cold-weather extrapolation. section 9.9
A practical coverage question
Before claiming that a system works broadly, ask which operating conditions its data actually represent. A tractor moving row by row through a predictable field needs less coverage than a vehicle navigating an uncontrolled city. The wider the claim, the more edge cases and conditions the data must cover. section 9.10 section 9.9
