Public reproducibility bundle for the paper. Self-contained and ready to push to a dedicated public repository.
This is the group-level companion to Personal Salience (glasp-co/personal-salience)
and Selection, Not Salience (glasp-co/personal-selection). Those papers showed the
individual's within-document signal is a whisper and that individuality lives in
selection; this paper asks whether the within-document crowd is internally structured
into reader sub-groups — and whether that structure is a stable reader trait.
Using Glasp (a social web highlighter) and margin-preserving permutation nulls (a curveball null that removes shared salience, mark density, and sentence popularity; and a stricter region-preserving null), we ask the group-level question and find:
- Experiment 1 — within a document, readers form STRONG sub-groups. Pairs agree with each other far beyond chance: nearest-neighbour agreement z = +6.3 (88% of documents significant; observed cosine 0.588 vs 0.466 null, excess +0.122); variance z = +12.8 (95%). The standardized effect is density-dependent (z 3.3-6.3); the raw decomposition below is stable. A loud group-level signal exactly where the individual signal was a whisper (+0.017, Paper I).
- …and we decompose it. Against an eight-block region-preserving null (which holds each reader's per-section engagement fixed), shared region engagement accounts for about 40% of the raw agreement excess (0.051 of 0.122); the majority (~60%, the 0.071 that survives region- matching, z = +3.6, 77% significant) is a finer reader-specific component (the position blocks only approximate true sections, so this finer share is an upper estimate).
- Experiment 2 — is that grouping a stable reader trait? UNRESOLVED. The cross-document split-half reproducibility is near zero pooled (+0.078 and 0.000 in two separately drawn samples). The test is informative only for pairs that co-read many documents. By shared-document count: k=2 +0.083, k=3 −0.002, informative k≥4 +0.220 [−0.17, 0.46] (and +0.34 [−0.39, 0.67], n=21, in the second sample). No stratum is significant; the high-k estimate is non-monotonic, imprecise across the separately drawn samples (the pooled second sample is 0.000), and attenuates under a region-preserving null (k≥4 +0.245 [−0.27, 0.54]). Every interval spans zero (all CIs in the paper's Experiment 2 table).
The within-document crowd is, in a descriptive sense, factional; whether its factions follow the reader across documents is, honestly, beyond the reach of the available co-readership density.
Experiment 1 (curveball null; 75 dense documents, by-document cluster bootstrap):
| statistic | mean z vs null | docs significant | | --------------------------------------------- | ------------------- | ---------------- | -------------------------------------- | | nearest-neighbour agreement (full null) | +6.3 [5.4, 7.3] | 66/75 (88%) | | pairwise-agreement variance (full null) | +12.8 [10.7, 15.1] | 71/75 (95%) | | nearest-neighbour, region-preserving null | +3.6 | 54/70 (77%) | (~60% of excess survives; region ~40%) |
Experiment 2 (cross-document split-half; 744 pairs, by-pair / by-reader bootstrap):
| shared documents k | pairs | split-half | 95% CI |
|---|---|---|---|
| k = 2 (no power) | 555 | +0.083 | [−0.03, 0.20] |
| k = 3 | 100 | −0.002 | [−0.14, 0.13] |
| k ≥ 4 (informative) | 89 | +0.220 | [−0.17, 0.46] |
| pooled k ≥ 2 | 744 | +0.078 | [−0.02, 0.18] |
Calibration anchors (synthetic ground truth): realistic-stable 0.15, strong-stable 0.43. No stratum is significant; stability is unresolved.
paper.tex— the paper (compile with pdflatex / arXiv).paper.pdf— the compiled paper (11 pages, 3 figures, 3 tables).figures/— three figures (PDF + PNG): the conceptual factions diagram, the per-document structure z-distribution against both nulls (E1), and the k-stratified cross-document stability with calibration anchors (E2).
Every effect size carries a cluster-bootstrap confidence interval, and both tests are validated on synthetic data with known ground truth before being trusted on real data (a no-structure control, a planted-group control, and a stable-vs-situational control for E2). The data-extraction and scoring pipeline runs against Glasp's private user data and is not released; per-pair results derive from individual highlighting behaviour and are therefore not published. Aggregate statistics and the estimator are available to researchers on reasonable request.
@misc{nakayashiki2026factions,
title = {Factions Within, Uncertain Across: Within-Document Reader
Sub-Groups in Social Highlighting},
author = {Nakayashiki, Kazuki and Watanabe, Keisuke},
year = {2026},
eprint = {2606.11613},
archivePrefix = {arXiv},
primaryClass = {cs.IR}
}Paper (paper.tex, paper.pdf) and figures: CC BY 4.0. See LICENSE. No code or data is
released in this bundle.