Progressive disclosure: showing a tree of text nested to any depth

Question: how is a tree of units nested to any depth shown to a reader, so the top levels are kept and any leaf is reached? And what does the evidence say about depth, expansion and navigation in nested text?

Files

File Role
protocol how the pool on folding, zooming and hypertext was searched and scored
sources the flow from hits to sources, then each source with its quality
frame step 1: the problem, and what interface, hypertext and model research call it
review step 2: fisheye views, folded text, hypertext structure, outliners, summary trees
state of the art step 3: most adopted against best evidence; what a capped view can show
position step 4: depth in storage against depth in view, and its three conditions
against readers who leave folds closed, null results, the limits of the studies
baseline step 5: three heading levels with contents beside, and document-order retrieval
bibliography every cited work, shared by the three reviews
flow.json the flow as data, drawn in the figure of sources

The repository's question

  • Can depth be unbounded in storage and capped in what a reader sees? The evidence allows it as a design, supported with caveat, never measured as such.
  • Three conditions come with it: show the path to the root, keep what most readers need above the first fold, and give every folded node a summary that stands for it (position).

Claims

Claim Grade Where
A fisheye of a tree keeps the focus, its neighbours and its ancestral line; in a regular tree its size grows with the logarithm of the tree supported with caveat (a property of the definition, Furnas 1986) review.md 1
Fisheye views beat flat windows for judging order in an unfamiliar hierarchy: 52% correct with two flat views, 64% with one fisheye, 75% with two measured, one study, 20 subjects review.md 1
Expand/contract contents beat a fully expanded list for retrieval in a three-level tree of 1,296 items on five of nine tasks, lost on the easiest, and did not differ on three measured, one study, 41 novices review.md 2
Overview plus detail gives the best essays about a document; shrinking text by importance is faster, with grades no better than linear text and less incidental learning measured, one study, 20 subjects review.md 2
Readers of a layered document answer less accurately on topics left out of its first layer (measured); the authors infer that few went past it measured, one study, 749 Internet users; the inference not measured review.md 2
Sections open on load drew more reading time than folded ones on mobile Wikipedia convention, with data: one uncontrolled test, on a grey source review.md 7
Hierarchical links beat network links for novices finding facts; hierarchy with cross-links can beat both measured, two studies, 64 and 30 subjects review.md 3
No interface for focus and context wins on every task; distortion rarely improves performance measured (one review) review.md 0; against.md
On maps and trees, readers preferred an overview without performing better with it measured, two studies (both secondary) against.md
For strong long-context models, passages in document order match or beat trees of summaries measured (benchmarks, matched budgets) review.md 6
For agents, one level of disclosure is enough, and a second can collapse accuracy measured, one preprint review.md 6
A capped view must elide the ancestors of nodes deeper than the cap inference from the definitions state-of-the-art.md
Outliners store and show unbounded depth and are widely used: 1,471,149 downloads of one plugin, 45,407 stars for one tool measured, descriptive state-of-the-art.md
At most two disclosure levels; never an accordion inside another convention, untested review.md 5, 7
Depth unbounded in storage, capped at three levels in view, with the path to the root shown supported with caveat (no direct test) position.md

Named baseline

  • Human reader: B1, the same content as one linear document with at most three heading levels and its contents always visible beside it (overview plus detail); B0, the fully expanded tree, as the control.
  • Model reader: DOS RAG, passages retrieved in the document's original order at a matched token budget (Laitenberger 2025); one level of disclosure for agents (He 2026).
  • Comparison: timed retrieval at known depths, blind-graded summaries, recall by level, opens below the fold, path naming; crossover with a one-week delay.
  • The open question gets its own test: storage depth 3, 5 and 7 under a view capped at three (baseline).

General principles linked, not reviewed again