Most research on robot map-learning formulates the task as learning a complete map of the environment during a mapping phase, and then using the constructed representation for goal achievement. We suggest that this view of a mapper, as a system that outputs an essentially static representation, is misguided. Rather, a mapping system should be considered an ongoing resource for planning, but one which adapts to the environment it finds itself in. We have developed a framework for designing such adaptive models, in which we have designed a system for mobile-robot map learning. The world is modelled as a set of places with relative positions and known actions that get the robot from place to place. Adaptive modelling requires that the modelling system not need to control the robot, so we achieve passive mapping by correcting mapping errors after they occur. Even in this paradigm, the mapper can also give the planner advice on certain actions that would help with mapping; if the planner decides to execute them, mapping will be improved, but if not, no harm is done. We are currently developing this system in simulation; some results are presented.