Field questions
Navigation questions, resolved
Practical starting points for deciding what to automate, what to measure and what a failed run can honestly tell you.
What does automated navigation testing actually test?
It checks whether a specified agent completes declared journeys under controlled conditions. Planning, movement, interactions and arrival checks may need separate tests. Start with a route contract that states what is and is not being exercised.
Does a valid path mean the route is playable?
No. A valid abstract path may still fail because of collision, animation, timing, world state or locomotion constraints. Compare the planner’s assumptions with the character’s real movement capabilities.
Do I need a machine-learning agent?
No. Scripted agents often provide a clearer baseline for known routes. Use learned policies only when their behavior serves a defined testing purpose. Our agent comparison explains the trade-offs.
How should a test detect a stuck agent?
Combine route progress and displacement over a bounded window with awareness of expected waiting states. Set thresholds for the route and movement mode. A lift wait and a blocked doorway should not share an unqualified “no movement” verdict.
What makes a route reproducible?
Record the build, map, route, movement profile, relevant world state, harness configuration and seeds. Document uncontrolled nondeterminism. Preserve the original artifacts before creating a smaller reproduction.
How much navigation coverage is enough?
There is no universal percentage. Define a route inventory and prioritize progression-critical transitions, recent changes and complex system boundaries. Read Count meaningful journeys before choosing dashboard metrics.
Can automated tests replace human playtesting?
No. Automation checks repeatable contracts, while human playtesting evaluates readability, usability, enjoyment and unexpected interpretations. Both can reveal defects the other misses.
What should a failed run produce?
Keep the setup, explicit outcome, synchronized trace, spatial context and evidence limitations. Preserve the original failure before reducing it. The evidence guide provides conceptual trace notes.
Does a fixed random seed guarantee an identical replay?
No. Physics, scheduling, hardware and asynchronous systems can still change execution. Record those factors and repeat runs where necessary. A seed is one input, not a complete description of the environment.
Does the contact tool submit my message?
No. It creates a local plain-text preview that you can copy or download. You choose whether to send it separately by email. Open the route-brief tool to prepare a question.