Breadth across Siftly bookmark categories (out of 13) + GDELT news domains (out of 30), averaged. Kills single-community fads.
0.00 single niche · 0.30–0.50 several areas · 0.70+ pervasive
source_diversity
Do multiple systems see it?
Count of independent sources (Google Trends, Siftly, GDELT) divided by 3. Kills artifacts of any one channel.
0.33 one source · 0.67 two sources · 1.00 all three
A fifth signal — internal_correlation — was designed into the composite but never implemented. It was dropped on 2026-04-20 after 214-record analysis showed zero signal across the entire field.
Only tiers 2 and 3 fire a Telegram alert. Tier 2 means a topic is moving meaningfully. Tier 3 means it's a genuine cultural breakout — rare by design.
Where the topics come from
Sliver doesn't pick topics out of thin air. The candidate list each run is assembled from three sources in priority order, then handed to Google Trends for scoring:
1
OpenCLI trend monitor
Reads SLIVER_OPENCLI_TOPICS — a JSON array written by an external trend-detection cron. Dormant in production right now, so we fall through.
2
Siftly bookmark extraction active
Pulls the last 14 days of bookmarks from disco-cli.exe.xyz, regexes out Title-Case 2–3-word phrases ("Claude Code", "Raspberry Pi"), filters stopword starters, ranks each phrase by frequency × recency with a 72-hour half-life, returns the top 12.
3
Curated Google Trends fallback
10 hardcoded tech terms (artificial intelligence, machine learning, open source, cloud computing…) merged in after Siftly, deduped case-insensitively, capped at 12.
The implicit contract: your bookmarks are the product. A week of India/classical-arts reading produces a week of India/arts topics in the scoring field. The curated fallback is effectively dormant because dynamic candidates always fill the 12 slots. If you want different topics, bookmark different things — there's no LLM topic selection, no magic pool, no web crawl.
The Siftly connection
Sliver and disco-cli are a single system. Two independent connections, both live every run:
Candidate extraction
GET /api/bookmarks?limit=200
Pulls recent bookmarks → phrase-extracts the 12 candidate topics for this run.
Per-topic signal
GET /api/bookmarks?q=<topic>
For each scored topic, counts matching bookmarks (feeds source_diversity) and categories they span out of 13 (feeds cross_domain).
What happens when disco-cli goes down: Sliver keeps running, but falls back to the 10-topic curated list and loses 2 of 5 scoring signals. Scores collapse toward a TimesFM+velocity floor — no triangulation, no spread detection, no "is this in my bookmarks" signal. Effectively reverts Sliver to a first-generation trend scorer.
Known drag on the signal: Siftly's /api/categorize enrichment is broken for 147 of 681 bookmarks (no category slugs written), which is partly why cross_domain_score has median 0.000 in the 72-hour distribution. Fixing enrichment would mechanically lift the cross-domain signal across the whole field.
Live examples · today's run
The current topic field, scored under the new weights and ranked. Reading each row is the same exercise every time: look at the four bars, ask whether they triangulate or contradict each other.
Claude Code Alert
0.737
timesfm
0.86
velocity
0.30
cross_domain
0.85
source_diversity
1.00
All four lit up. High forecast, moderate growth, very wide spread, all three sources seeing it. The clearest "exploding topic" in the field — close to tier 3. Anthropic's CLI is eating the dev world in real time.
Open Source Watch
0.524
timesfm
0.48
velocity
0.06
cross_domain
0.77
source_diversity
1.00
Everywhere (high spread, every source), but not moving. Flat forecast, near-zero velocity. A background constant, not an event. Sliver surfaces it, but there's nothing to act on.
Raspberry Pi Watch
0.520
timesfm
0.73
velocity
0.06
cross_domain
0.39
source_diversity
1.00
Forecast says rising, velocity hasn't started yet. Moderate domain spread. This is the early pattern — before growth is obvious. The most actionable read: if you were going to build or write something Raspberry-Pi-adjacent, now would be the window.
Nataraja Temple Chidambaram Watch
0.481
timesfm
0.83
velocity
0.36
cross_domain
0.00
source_diversity
0.67
Rising forecast, real velocity, but zero cross-domain spread. Classic single-community shape — interest concentrated in one area (likely Indian classical / spiritual-travel), not bleeding into tech or news. Trust the signal if you're in that community; ignore otherwise.
Notebook LM Watch
0.458
timesfm
0.96
velocity
0.09
cross_domain
0.00
source_diversity
0.67
Forecast is at the ceiling (0.96) but velocity is near zero and cross-domain is zero. TimesFM is very sure it's rising, but the rise hasn't manifested in the other signals yet. Either the earliest possible early signal — or TimesFM is hallucinating a trend that isn't there. Watch this one resolve over the next 24hrs.
Golden Ratio Watch
0.425
timesfm
0.60
velocity
0.00
cross_domain
0.27
source_diversity
1.00
Three sources see it but no velocity. Evergreen topic — perpetually searched, never breaking out. Sliver's source_diversity is doing real work here: otherwise this would look flat-zero.
Natya Shastra Watch
0.403
timesfm
0.70
velocity
0.00
cross_domain
0.03
source_diversity
1.00
Same shape as Nataraja Temple. Forecast says up, velocity flat, single subject area. This is what a niche, slowly-rising cultural reference looks like in Sliver's field of view.
Matt Pocock Watch
0.345
timesfm
0.70
velocity
0.00
cross_domain
0.00
source_diversity
0.67
Just forecast and 2-source presence. No growth, no spread. A creator/personality with steady attention. If you care about him, it's fine signal — if not, it's legitimately quiet.
Geometric Analysis None
0.245
timesfm
0.20
velocity
0.00
cross_domain
0.10
source_diversity
1.00
Forecast is saying down (anything below 0.50 is negative). Nothing else firing. This is what a dying term looks like in Sliver — quietly fading, surfaced only by source_diversity accident.
Bharatanatyam Postures None
0.103
timesfm
0.01
velocity
0.00
cross_domain
0.00
source_diversity
0.67
Essentially flatlined. Forecast at the floor, no velocity, no spread. A candidate from a bookmark that the scoring layer has nothing to say about. Useful as a negative benchmark — this is the bottom of the distribution.
Reading any alert in one sentence
tf + velocity high, cd + sd low — single-community hype. Skip unless you're in that community.
tf high, velocity low — early, still forming. Best time to write / build.