Just as linguistic errors have provided insights into the nature of linguistic competence(s), detailed analyses of blind spots or "marked trajectories" in the reflection cycles of novice teachers may provide a window on their underlying beliefs, and thus on relevant zones of proximal development. In this paper we analyse a case study in an experimental web-based learning environment in which a novice teacher demonstrably just reproduces the assumptions she started out with rather than re-inspecting them. Having located a problem of motivation uniquely inside the learners' heads, she pointedly ignores hints in the feedback that her own verbal and nonverbal behaviour in the classroom might be a relevant domain of enquiry. Our findings show the need for structural interventions within "a priori" hypothesized, linearly ordered, stages in reflection models. We report on two small-scale experiments that implement suggested changes in the architecture of the web site which yield more context-sensitive ways of scaffolding reflection. In conclusion we argue that detailed discursive accounts of successful and less successful reflection trajectories are needed to refine and further develop models in teacher thinking. (Contains 4 figures and 2 notes.)