The intervention window must be long enough
Aliases: time to intervene · takeover margin · intervention time budget
What it is
A human intervention window runs from detectable abnormality to unavoidable or irreversibly unsafe action. It is real only if it covers detection, comprehension, choice, reaching control, system response, and physical cessation; a few logged milliseconds are not usable intervention.
Why it happens
Available time is hazard evolution minus sensing, display, human, communication, and braking delays. A late request at high speed makes supervision ceremonial. Earlier prediction, reduced speed, delayed commitment, and prepared safety state enlarge the window without demanding constant vigilance.
Studying it
Warning lead, speed, workload, and control complexity can vary while measuring response distributions, near misses, and completed braking. High quantiles rather than means represent fatigue, diverted attention, and accessibility. System delays should be measured separately from human delay.
Where it stops holding
Fast hazards need automatic protection. Longer windows with excessive false alarms reduce response, and unlimited slowing can make work infeasible. Expert laboratory response does not represent incidental bystanders.
Applying it
- Budget worst credible sensing, cueing, high-percentile human response, transmission, and physical stop.
- If insufficient, trigger earlier, limit speed, or use automatic protection rather than nominal supervision.
- Drill attention diversion and network variation, comparing the window with completed intervention and auditing every expired request.