Public reproducibility bundle for the paper. Self-contained and ready to push to a dedicated public repository.
This is the fifth paper in a program on social highlighting, after Personal Salience (arXiv:2606.09024), Selection, Not Salience (arXiv:2606.10398), Factions Within, Uncertain Across (arXiv:2606.11613), and The Long Tail, Not the Front Page (arXiv:2606.11654). Those papers located individuality in selection (which documents a person highlights) and measured it cross-sectionally; this paper adds the time axis: is a reader's selection signature a trait or a state?
Freeze each reader's first six months of highlighting as a profile; measure its own-vs-other advantage on the reader's later selections at growing gaps (up to 24+ months), against time-matched negatives (same calendar era, so supply drift cannot masquerade as personal drift), at two difficulty layers (global vs the reader's own interest neighborhood).
- The anchor replicates: fine-layer advantage +0.188 [+0.160, +0.216] right after the profile freezes, matching the prior cross-sectional estimate (+0.169).
- No detectable decay (the headline): within the same users, paired retention at a 6–12 month gap is R = 1.00 [0.85, 1.18] (n=212; declines larger than ~15% are excluded); no horizon up to 24+ months shows a significant paired decline in the fine layer (the farthest bin, n=65, remains compatible with a modest late decline). The only significant decay anywhere: the coarse topic layer at 12–24 months (−0.067, ≈13%).
- Not reducible to repeated domains: on matched cells, removing every profile domain retains 91–96% of the advantage at every horizon (ratio CIs ~0.84–1.02).
- Slow drift within the durable identity: volume-matched, the recent half of history beats the old half by +0.042 [+0.020, +0.064] — yet the old half retains ~91%.
- Prospectively actionable: every personal variant — even a profile from the reader's earliest documents (0.66 AP, ~94% of the recent profile) — beats every non-personal prior by ~3x (lifetime popularity 0.23 ties with random; neighborhood co-reading 0.19 falls below it).
Honest scope: heavy, long-tenured readers of one platform (1-in-472 of scanned user records); exposure vs choice is not separable without impression logs; the instrument is a frozen embedding centroid (a lower bound, compared with itself across gaps).
| Result | Value |
|---|---|
| anchor advantage (fine layer, 0–1 mo) | +0.188 [+0.160, +0.216] (prior cross-sectional: +0.169) |
| paired retention R, 6–12 mo (primary) | 1.003 [0.854, 1.184], d = +0.001 [−0.031, +0.032], n=212 |
| paired d, 12–24 mo / 24+ mo (fine) | −0.019 [−0.054, +0.016] / −0.031 [−0.078, +0.016] (both n.s.) |
| only significant decay (coarse, 12–24 mo) | −0.067 [−0.108, −0.026] |
| held-out (no profile domains) share retained | ≥83% at every horizon (83→93%) |
| recent half − old half (within-user, volume-matched) | +0.042 [+0.020, +0.064], n=248 |
| next-doc: personal / popularity / random | 0.704 / 0.229 / 0.229 AP (gate +0.475, pass) |
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 identity decay curve, the within-user paired retention panel, and the slow-drift + prospective-gate panel.
Every effect size carries a 3,000-iteration by-user cluster-bootstrap CI; the cohort definition, profile construction, gap bins, regimes, and pass rules were fixed in a versioned design document before each run (the two-regime amendment was registered after a 60-user smoke test and before the full-cohort run, and is disclosed in the paper). The data-extraction and scoring pipeline runs against Glasp's private user data and is not released; per-user results derive from individual highlighting behavior and are not published. The cluster-bootstrap estimator and aggregate statistics are available to researchers on reasonable request.
@misc{nakayashiki2026trait,
title = {Trait, Not State: The Durability of Reading Identity
in Social Highlighting},
author = {Nakayashiki, Kazuki and Watanabe, Keisuke},
year = {2026},
eprint = {2606.12904},
archivePrefix = {arXiv},
primaryClass = {cs.IR}
}Paper (paper.tex, paper.pdf) and figures: CC BY 4.0. See LICENSE.