In reply to Defining notan's edge:
[John] letters have a built-in advantage over sub-letter and super-letter elements as perceptual units in that they are self-contained visually and often linguistically (more so in many languages than in English), and b) the optimal perceptual unit varies as we read, most obviously relative to word length and familiarity.
John, think in perceptual processing terms. To facilitate this I suggest it is important to distinguish between units of perception and units of processing, tokenizable parts of composite structures, and featural primitives.
Words have a built-in advantage over letters as perceptual units in that they are space-delimited and internally cohesive bounded maps of blacks and whites. To get to letters a preliminary segmentation (tokenization) must occur within this unit that designer try their darndest to make internally cohesive, that is, to make into a single unit, a perceptual unit.* Letters are the tokenizable parts of composite structures and the composite structures have gestalt integrity.
Every perceptual scientist I know begins with feature detection or feature analytic processing. To me features (in the alphabetic domain) are things like expressedness and closure. In my scheme the units of processing in glyph-like structures are role-units. Most psychologist collapse the two (features and role-units).
If, at the role-unit level, blacks are in phase and whites rhythmically coordinated, the segmentation is psycho-physically discourged or inhibited. This is why I think a single-tiered cross-letter gathering of role-units (protoypical structures — white and black — not literal shapes) with attention to their local combination characteristics and global distribution across word-integral between-letter reference points is the best way to address the recognitional problem.
* Should designers stop trying to do this? Don't designers try to acheive gestalt cohesion by making letters less self-contained and more open to eachother?