Often features are presented as clean, neat, simple. Indeed it is the contrast with the idiosyncrasies of lexical items which gives the intuitive justification for features. But reality is more complex. There are many instances where it is arguable whether we should postulate a feature (value), as with person in Archi. We must recognize that feature systems vary: (a) according to how well founded they are, and (b) in how they distribute across the lexicon. To analyse this difficult area, the penumbra of feature systems, I start from an idealized view, and then plot the deviations from that ideal. In other words, I take a ‘canonical’ approach. Having justified this approach in general terms, I propose a specific set of converging criteria for canonical features and values, concentrating on the genuine morphosyntactic features. In brief, the overarching principles are that a canonical morphosyntactic feature is constrained by simple rules of syntax (including the claim that syntax is ‘morphology-free’) and has robust formal marking. These give us a point in the theoretical space from which to calibrate the difficult instances which abound in feature systems. In accounts of particular features, various types of what we may call non-canonical behaviour have been pointed out: e.g., non-autonomous case values (Zaliznjak 1973), minor numbers, inquorate genders. We should ask whether these problems are feature-specific or whether they recur in the different morphosyntactic features. It turns out that, at the right level of abstraction, we find similar instances of non-canonicity with the different features. Let us concentrate on the criteria contributing to ‘robust formal marking’: Criterion 1: Canonical features and their values have dedicated forms. We find non-autonomous case values (violating criterion 1), in Classical Armenian, for instance (Baerman 2002); similarly we find non-autonomous gender values (as in Romanian). Criterion 2: Canonical features and their values are uniquely distinguishable across other logically compatible features and their values. Deviations give sub-genders (Serbian/Croatian/Bosnian), sub-cases (Russian) and sub-numbers (Biak). Criterion 3: Canonical features and their values are distinguished consistently across relevant parts of speech (word classes). In the easy examples, one part of speech has values which represent a collapsing of values available on another. More interesting are systems where combinations give additional values: combined gender systems (Mba), constructed number systems (Mele-Fila) and combined person (Maybrat). Criterion 4: Canonical features and their values are distinguished consistently across lexemes within relevant parts of speech. The basic deviation gives us a minor value: as in a minor number value (Bezhta), a minor case (Russian), a minor gender (Lelemi). This leads to the question of whether such deviations can co-exist. I give a striking example: the Russian second genitive exhibits all four types of non-canonical behaviour concurrently. Since it deviates from all four listed criteria this marks the extreme of the typological space. By investigating the penumbra of feature systems, including the possible and impossible interactions within the penumbra (which features are the context for the deviations of others), we put the theory of morphosyntactic features on a more realistic and hence firmer foundation.