Find posts whose bookmarks
are starting to grow.
Elontir helps you discover adult posts on X through bookmark growth and account connections.
@mia_nanasawa 90.0Bookmarks/hour 126Bookmarks 114Reposts 1 hour ago 
@mayukiito 44.9Bookmarks/hour 64Bookmarks 61Reposts 1 hour ago 
@ako_kimu 86.0Bookmarks/hour 86Bookmarks 56Reposts 57 minutes ago 
@yura_kana_ 29.4Bookmarks/hour 257Bookmarks 135Reposts 9 hours ago 
@Aizawa_miyu03 11.0Bookmarks/hour 11Bookmarks 11Reposts 23 minutes ago 
@yaginana0903 106Bookmarks/hour 2.1KBookmarks 256Reposts 20 hours ago Showing the highest early-momentum scores among 199 posts observed while monitoring 60 accounts. Tap an image to open it on X. Bookmark pace is the hourly average since posting. View the latest rankings
How posts are discovered
Why cumulative rankings get stuck
Most bookmark rankings sort by cumulative totals. Older posts have had more time to build those totals, so works that gain momentum later cannot reach the top.
Sort the same posts by two different criteria and the lineup changes. The lists below use real data from this observation. The list on the right includes posts absent from the cumulative ranking.
Ranked by total bookmarks
- 3.9K @ako_kimu
- 2.1K @yaginana0903
- 2K @rena_miyashita
- 1.7K @mia_nanasawa
- 1.7K @ako_kimu
Ranked by bookmark growth rate
- 106/h @yaginana0903
- 90.0/h @mia_nanasawa
- 86.0/h @ako_kimu
- 81.5/h @ako_kimu
- 55.2/h @ako_kimu
A post with a bookmark count of 126 (@mia_nanasawa) ranks near the top by velocity. It does not appear in the cumulative list.
Find posts through account connections
Start from one account and follow connections to expand the search. We use three kinds of connections: follows, reposts, and related-account candidates.
- Explore
Follow connections from a starting point
Follow nearby connections first. Accounts reached through multiple routes are treated as more relevant and prioritized. One starting point expanded into 1,495 accounts.
- Filter
Retain relevant accounts
Accounts with the most followers include companies, local governments, and games. An LLM classifies relevance, retaining 1,183 out of 1,495. Classification takes 1.5 seconds per account.
- Measure
Measure velocity and growth
Record bookmarks, reposts, and likes separately, then measure changes by observing again every 30 minutes. Dividing totals by elapsed time gives an average and misses the onset of growth.
- Discover
Bring new faces to the top
Identify early momentum in posts that have few bookmarks but a high growth rate. These are discoveries that cumulative rankings do not surface.
Exploration coverage
Connections actually followed from one starting account. Dots are accounts; lines are follows. Orange dots are strongly related accounts reached through multiple routes. Hover over a dot for details.
- Accounts explored
- 1,495
- Identified as industry-related
- 1,183
- Connections between accounts
- 1,914
- Posts collected
- 199 posts
- Reached through one route
- Multiple routes
- Four or more routes
FANZAdougaX
Up to 50 levels deep
At each account, follow the top 50 accounts by follower count and explore beyond the deepest level each cycle.
1,495 accounts
1,183 accounts
79% passed classification. The rest were excluded as companies, local governments, or unrelated accounts.
1,914
431 accounts
Accounts reached through two or more routes, shown in a different color on the map.
Experimental live
An experiment feeding post features into a fly-brain model to observe its responses. This is not a video stream of the posts and does not evaluate preferences or quality.
Open experimental live · Sort by experimental score
Model and input details
We continuously run a whole-brain simulation of an actual connectome (MaleCNS v1.0, 165,733 neurons / 25.59 million synapses). Features of posts gaining momentum are fed into the olfactory circuit as “odors,” and the responses of the memory center and value-assessment circuits are displayed.
Connectome: MaleCNS v1.0 (CC-BY 4.0, HHMI Janelia / Cambridge / MRC LMB / Google Research). This is a designed mapping for demonstration purposes and does not assign biological meaning. Read the full Creative Commons BY 4.0 license
The hero video is an edited recording of this simulation, cropped, recolored, and retimed.
The pulsing connection graph in the background is decorative and is not a live measurement.
Experimental-score validation data
Brain responses and actual growth
We compared simulated brain-response strength (R) with actual bookmark growth across 153 posts with observations. Bookmark increases were confirmed on 91 posts.
How to read this
- The top 25% by R had about 10.3 times the bookmark growth of the bottom 25%
- Posts with stronger responses tend to gain more bookmarks. However, this correlation comes from a designed mapping of post features to stimulus intensity and does not establish causation.
- Reposts correlate weakly with R (ρ=0.229)
- Reposts tracked the response more closely than bookmark growth (ρ=0.174). This may be because spreading activity tends to align more closely in time with the initial response to a stimulus.
- What is this for?
- This score is available when sorting rankings by experimental score. These values have meaning only within the simulation; it does not measure users’ psychology.
Measured collection and classification data
What the measurements confirmed
Only numbers obtained by running the system are shown here, not estimates.
Measure growth for each signal
Bookmarks, reposts, and likes behave differently. Equal counts do not mean the same thing, so we measure them separately and assign different weights.
- Reposts move first
- Over a 22-minute observation, reposts grew at a rate of 66 per hour. The same post gained no bookmarks. Reposts are the fastest indicator of emerging momentum.
- Spreading and saving are different
- In the same 22 minutes, one post had no bookmark growth and 66 reposts/h, while another had 5.5 bookmarks/h and 30 reposts/h. We read the first as spreading and the second as support.
- Use observed changes, not averages
- Growth is calculated from the change since the previous observation. Dividing totals by elapsed time gives an average, which misses the moment growth begins.
Flaws found during implementation, and the fixes
- Velocity is overstated just after posting
- One bookmark six minutes after posting was calculated as 10 per hour, placing the post above others that were actually growing. A minimum denominator for velocity fixed this.
- Greeting posts take first place
- Posts without video or images topped the velocity ranking. We added a media filter and now exclude posts without media by default.
- Classification returns an empty result
- Models that spend tokens on reasoning can return an empty response when the limit is too low. We raised the limit and added a fallback that recovers the result from the reasoning log.
About the person building this
This system grew out of experience running a bookmark-ranking site. That site relied on other sites’ aggregates, so it stopped working when they introduced countermeasures. The numbers were not our own, either.
We are solving the same problem again, this time with primary data: taking bookmark counts directly from X posts, discovering them through account connections, and ranking them by velocity. The measurements accumulate locally.
- Experience operating bookmark-ranking media
- Experience operating manga media across multiple domains
- Hands-on testing and reporting on cameras, video, and software
- Automating collection, classification, and monitoring with AI agents