Baseline: what the node template is compared against
This file names the baseline a nested, foldable node must beat, for a human reader and for a model reader, and how the comparison would be measured. Terms are defined in frame.
The named baselines
| Reader | Baseline | Why this one |
|---|---|---|
| Human | B1, shallow outline with overview plus detail: the same content as one linear document, at most three heading levels, with its contents list always visible beside the text | overview plus detail gave the best essays and was preferred in the one document-reading comparison (Hornbæk 2003, p. 140); it is the standard practice (Cockburn 2009); the three-level cap is the shallow outline the node must obey anyway |
| Human, control | B0, fully expanded tree: every level open, no contents pane | the "stable" interface of Chimera 1994, which isolates the effect of folding |
| Model | DOS RAG: passages retrieved and kept in the document's original order, at a matched token budget (Laitenberger 2025) | it matched or beat summary trees and an agent on three benchmarks |
| Model, agent | one level of disclosure: a short description, then an index of chunks, read on demand (He 2026) | no cell rewarded a second level: it tied one level within noise or trailed it (He 2026, tables 1 and 2) |
- The node must beat B1, not B0: beating a fully expanded tree is already shown (Chimera 1994), and B1 is what a careful writer would do without the node.
- For models, beating RAPTOR (Sarthi 2024) is not enough, since simpler baselines already do (Laitenberger 2025).
Candidate variant to test
- Each node opens with a summary that stands for it when folded.
- A view shows at most three levels: the focus, its children and its siblings, with the path to the root always visible.
- The reader can zoom into any node, which then becomes the root of the view.
- What most readers need sits above the first fold.
- These are hypotheses, graded convention, untested.
How a comparison would be measured
| Reader task | Outcome | Instrument | Prior study to borrow from |
|---|---|---|---|
| Find a fact | time to correct answer; accuracy | timed questions whose answers sit at known depths: top, middle, leaf | Chimera 1994 (nine timed tasks) |
| Understand the whole | blind-graded summary or essay; recall of high-level against low-level idea units | essay after reading; free recall scored by level | Hornbæk 2003; the levels effect (psycholinguistics) |
| Reach below the fold | share of readers opening folded nodes; accuracy on questions answered only below the fold | logged opens; questions placed by depth | McDonald 2009; Wikimedia 2016 |
| Keep bearings | correct naming of the path from the focus to the root; a lostness measure | prompt at random moments; navigation logs | the disorientation literature (Otter 2000, title only) |
| Depth in storage, with view capped | all outcomes above | the same content stored at depth 3, 5 and 7, each shown capped at three levels, against B1 | none found: this is the gap (position.md) |
- Design: within subjects, crossover, order counterbalanced, a delayed test one week later.
- Report effect sizes with confidence intervals; record preference, but never as the outcome, since preference and performance part ways (Hornbæk 2002).
- Subjects: novices and readers who know the domain, analysed apart: knowledge helped overall, and interacted with structure on nodes opened while reading (McDonald 1998, p. 22).
For a model reader:
- Benchmarks: QASPER, NarrativeQA and QuALITY, as in Sarthi 2024; ∞Bench, as in Laitenberger 2025 and He 2026.
- Conditions: node-tree navigation, DOS RAG, and one-level disclosure, at matched token budgets.
- Readers of at least two sizes, since gains from compression shrink with stronger readers (Laitenberger 2025).
- Outcomes: accuracy, tokens read, and whether the answer cites the right leaf.
A checklist for a document in the baseline's own terms
Per document, B1:
- At most three heading levels.
- Every heading names its content and stands alone when read out of context (WSE 2013, an opinion).
- The main claims sit at the top level, as headings or the first sentence of each part.
- A contents list is visible beside the text and links every section.
- Nothing is folded; whatever is optional is marked, not hidden.