You know where things are. You just cannot get to them.
You have a system. Notes, files, bookmarks, project folders, a knowledge base — something with structure. And you know, in a vague sense, that the thing you need is in there somewhere. The problem is not that the information is missing. The problem is that the path from "I need something" to "here it is" requires you to hold too much in your head at once.
This is the navigation problem. And it does not live at the top of your hierarchy or the bottom. It lives in the middle — in the layers between "everything" and "the specific thing." Those intermediate levels are where most hierarchies succeed or fail. Not because the root categories are wrong, and not because the leaf-level items are poorly named, but because the middle either guides you through or abandons you to scroll, scan, and guess.
The previous lesson established that root concepts anchor everything beneath them. Get the root wrong, and every subordinate category inherits the error. But even with perfect root concepts, a hierarchy can still be unusable if the space between the top and the bottom is empty, cluttered, or organized around the wrong questions.
The middle is where wayfinding happens
Kevin Lynch spent five years studying how residents of Boston, Los Angeles, and Jersey City navigate their cities. His 1960 book The Image of the City identified five elements people use to build mental maps: paths, edges, districts, nodes, and landmarks. The critical insight was not about any single element but about how they work together at different scales. A landmark orients you from a distance. A district tells you what neighborhood you are in. A node marks the point where you make a decision about which path to take next.
Cities that are easy to navigate have clear intermediate structure. You do not go from "I am in Boston" directly to "I am at 47 Elm Street." You go from Boston to Back Bay to Newbury Street to the block near the park to the specific address. Each step narrows your location without requiring you to process every possible destination at once. Lynch called this quality "imageability" — the degree to which an environment produces a clear, coherent mental image that supports navigation.
Hierarchies in your knowledge system work the same way. The root level is the city. The leaf level is the address. The intermediate levels are the districts, the neighborhoods, the streets — the structure that makes it possible to progressively narrow from "everything" to "this one thing" without cognitive overload.
When the intermediate levels are missing, you get the navigational equivalent of being dropped into a featureless grid: you know the address exists, but every block looks the same, and you have no landmarks, no districts, no neighborhoods to orient by.
What libraries figured out two centuries ago
Melvil Dewey published his decimal classification in 1876, and the core structural insight still holds. The system uses three levels encoded in a single number: the first digit identifies one of ten main classes (000-999), the second digit identifies one of ten divisions within that class, and the third digit identifies one of ten sections within that division. Decimals after the point allow further specificity.
The power is not in the numbering scheme itself. It is in what the divisions do for the person standing in the stacks. You want a book on American poetry. You do not need to scan every book in the library. You go to the 800s (Literature), then to 810 (American Literature in English), then to 811 (American Poetry). Three navigation decisions, each reducing your search space by roughly an order of magnitude.
The divisions — the hundreds-level categories — are the intermediate layer. They exist not because they contain information but because they help you find information. Nobody reads the 810 sign and thinks "ah, now I understand American literature." They read it and think "I am getting closer." That is the function of intermediate levels: not to be destinations, but to be waypoints.
This is what Rosenfeld, Morville, and Arango describe in Information Architecture: For the Web and Beyond as the navigation function of taxonomy. Classification systems serve two purposes: they organize content for storage, and they guide users during retrieval. These purposes can conflict. The way an archivist would organize a collection (by provenance, by date, by format) is often different from the way a researcher would search it (by topic, by question, by relevance to a current project). Intermediate levels work when they are designed for retrieval, not just for storage.
The cognitive math of progressive narrowing
George Miller's 1956 paper established that humans can hold roughly seven items (plus or minus two) in short-term memory. Nelson Cowan's subsequent research lowered that estimate to about four chunks when you strip away rehearsal strategies. The practical implication for hierarchy design: at any given navigation step, if you present more than about seven to ten options, you exceed what a person can scan, compare, and choose between effectively.
This creates a design constraint. If you have 1,000 items at the leaf level and you want each navigation step to present at most 10 options, you need at least three levels: 10 choices at the top, 10 at the middle, 10 at the bottom gives you 1,000 reachable leaves. Two levels would require 32 options at each step (the square root of 1,000) — past the point where scanning is comfortable. Four levels would let you reduce to about 6 options per step, but each additional level adds a click, a decision, and a chance to take a wrong turn.
Usability research on website navigation converges on a consistent finding: structures that are moderately broad at each level and moderately deep in total levels produce the best navigation performance. Very flat structures (all items at one level) overwhelm users with choices. Very deep structures (many levels with few items each) force users through too many sequential decisions and make it hard to recover from wrong turns. The optimal shape, according to human factors research, is slightly concave — broader at the top and bottom, narrower in the middle — giving users a wide entry point, focused intermediate navigation, and a rich terminal selection.
The intermediate levels are what make this concave shape possible. They are the narrowing point — the place where a broad initial category resolves into something specific enough to act on.
Your brain already does this
Deep neural networks learn hierarchical representations that mirror this same structure. Early layers detect low-level features — edges, textures, simple patterns. Final layers produce high-level outputs — "this is a cat," "this is a stop sign." But the intermediate layers are where the real work happens. They combine low-level features into progressively more abstract representations: edges become shapes, shapes become parts, parts become objects.
Research on feature visualization in convolutional neural networks shows that intermediate layers encode mid-level concepts that are neither raw input nor final classification. A middle layer might represent "fur texture" or "circular shape" — abstractions that are useful for navigation between raw pixels and named categories but are not themselves the answer to any specific question.
This is not a metaphor. It is the same structural principle operating in a different substrate. Intermediate representations exist because the gap between raw input and meaningful output is too large to cross in a single step. The middle layers do not contain the answer. They contain the progressive narrowing that makes the answer reachable.
Gestalt psychology identified the same principle in human perception. The law of proximity says we group nearby elements together. The law of similarity says we group similar elements together. These groupings are themselves intermediate representations — perceptual middle layers that reduce a visual field of thousands of individual elements into a manageable set of clusters you can attend to, compare, and navigate between. Without these automatic groupings, you would face the raw complexity of every individual stimulus at once. The Gestalt principles are your perceptual system's intermediate navigation layer.
Why organizations have middle managers
The same principle explains why flat organizations struggle past a certain size. A single leader with 100 direct reports cannot effectively coordinate, contextualize, or route information for all of them. Not because the leader is incompetent, but because the span of control exceeds what one person can navigate.
V. A. Graicunas formalized this in 1933, showing that the number of relationships a manager must maintain grows exponentially with the number of direct reports. With 5 reports, a manager handles 100 potential relationships. With 10, it is over 5,000. Middle management exists to create intermediate navigation layers — translating between the strategic abstraction of executive leadership and the operational specificity of individual contributors.
McKinsey's research on span of control confirms that the optimal number of direct reports depends on context but consistently falls in the range where intermediate layers are necessary for any organization beyond a small team. The middle layers are not bureaucratic overhead. They are navigational infrastructure. They exist so that information, decisions, and context can flow between the top and bottom of the hierarchy without either level being overwhelmed.
When organizations "flatten" by removing middle management entirely, they often discover that the navigation function does not disappear — it just becomes informal, inconsistent, and invisible. People create ad hoc intermediate layers through Slack channels, informal team leads, and tribal knowledge about "who to ask about what." The structure reasserts itself because the cognitive need for progressive narrowing does not go away when you remove the formal layer that was serving it.
AI as a dynamic intermediate layer
Here is where this becomes directly relevant to building your cognitive infrastructure. Traditional hierarchies have fixed intermediate levels — the folder structure, the category taxonomy, the org chart. You design them once, and they stay until someone reorganizes. This means the intermediate layer reflects the mental model of the person who built it, which may not match the mental model of the person searching.
AI changes this. A well-configured AI assistant can act as a dynamic intermediate layer — one that generates navigation structure on the fly based on the specific query, the context of the search, and the known structure of your knowledge base. Instead of navigating through fixed folders, you describe what you need, and the AI constructs an intermediate path: "Based on your question, the most relevant cluster is these 8 notes across 3 topics. Here they are, ranked by relevance."
This does not eliminate the need for well-designed static hierarchy. You still need good root concepts (L-0268) and well-named leaf items. But it supplements fixed intermediate layers with adaptive ones — navigation that reshapes itself based on how you are searching right now rather than how someone organized things months ago.
The key design principle: build your static intermediate layers for the most common navigation patterns, and use AI-assisted search for everything else. The static layers handle the 80% of searches that follow predictable paths. The AI handles the 20% where the fixed structure does not match your current question.
Protocol: designing intermediate levels that actually navigate
When you build or evaluate any hierarchical structure, apply these tests to the intermediate levels:
The narrowing test. At each intermediate level, does the number of options stay within a scannable range (roughly 5 to 12 items)? If a middle level presents 30 subcategories, it is not doing its navigation job. Split it into two levels or regroup.
The question test. Can you state each intermediate level as a question the searcher would actually ask? "What domain?" leads to "What topic within that domain?" leads to "What specific item?" If you cannot phrase the intermediate level as a natural question, the level is organized for storage rather than retrieval.
The wrong-turn test. If someone picks the wrong option at an intermediate level, can they recover quickly? Good intermediate layers have enough internal coherence that a wrong choice is obviously wrong within one or two steps. Bad intermediate layers create ambiguous boundaries where an item could plausibly belong in multiple categories, and you have to search all of them.
The retrieval-time test. Would the person searching for something six months from now navigate the same way the person organizing it today would? If the answer is "probably not," the intermediate layer is optimized for the organizer's current context rather than the searcher's future context.
The bridge to depth
Intermediate levels solve the navigation problem. But they introduce a new one: every level you add increases the total depth of the hierarchy. And depth has its own cognitive costs — each additional level means another decision, another click, another chance to lose your place or take a wrong turn.
The next lesson examines that tradeoff directly. L-0270 explores why flat structures are often preferable to deep ones — and how to recognize when you have added intermediate levels that help navigation versus intermediate levels that just add steps.
The principle from this lesson carries forward: intermediate levels exist to serve navigation, not to mirror the structure of the content. If an intermediate level does not help someone get closer to what they need, it is not navigation infrastructure. It is clutter.
Sources:
- Lynch, Kevin. The Image of the City. MIT Press, 1960.
- Rosenfeld, Louis, Peter Morville, and Jorge Arango. Information Architecture: For the Web and Beyond. 4th ed. O'Reilly Media, 2015.
- Miller, George A. "The Magical Number Seven, Plus or Minus Two." Psychological Review 63, no. 2 (1956): 81-97.
- Cowan, Nelson. "The Magical Number 4 in Short-Term Memory: A Reconsideration of Mental Storage Capacity." Behavioral and Brain Sciences 24 (2001): 87-114.
- Graicunas, V. A. "Relationship in Organization." Papers on the Science of Administration. Columbia University, 1937. Originally published 1933.
- Dewey, Melvil. Dewey Decimal Classification and Relative Index. OCLC, editions 1876-present.
- Wagemans, Johan, et al. "A Century of Gestalt Psychology in Visual Perception: I. Perceptual Grouping and Figure-Ground Organization." Psychological Bulletin 138, no. 6 (2012): 1172-1217.
- Olah, Chris, et al. "Feature Visualization." Distill (2017). https://distill.pub/2017/feature-visualization/