Every social platform eventually faces the same question: who is actually making this place better?
The loud answers are easy to find. Follower counts. Vote totals. Payouts. On Hive the question gets harder, because stake bends every one of those signals — one whale vote can outweigh a hundred humans, a tight circle of alts can manufacture the appearance of a thriving community, and the person quietly welcoming every newcomer shows up nowhere in the numbers.
We run 
We call it LNV — liketu network value. It has been computing daily in production since late June: every run scores the full cohort of liketu creators (386 accounts in the latest run) across a 90-day window, over a real interaction graph of 89,872 weighted edges built from on-chain votes, replies, and realized HBD. This post is about how it works, how it steers our voting, and why we think it points at something bigger than curation.
value lives in the graph
The founding observation is simple: on a social network, value is not a property of a person — it's a property of their connections. Attention, appreciation, and money all flow along edges: a vote, a reply, an HBD payout, a new user showing up because someone brought them.
This is not an illustration. It's the actual top of the liketu graph from the July 1 scoring run — every line is a real relationship made of replies, votes, and HBD flowing between two creators over 90 days. Thicker lines carry more realized value.
Once you see the network this way, the right measure of a person stops being "how big are their numbers" and becomes "how much genuine flow passes through them, from how many independent directions?"
That framing comes with a classic tool: eigenvector centrality — the same family of math as Google's PageRank. The intuition fits in one sentence: you matter when people who matter engage with you. A reply from someone the network genuinely values counts for more than ten drive-by votes from empty accounts, and the definition is beautifully recursive — their value was earned the same way. We compute it with a sparse power iteration over the edge graph, where each edge is weighted by what actually moved across it: realized HBD, distinct replies, distinct posts voted.
the anatomy of the score
Centrality alone isn't enough — it tells you about position, not behavior. So LNV is a weighted blend of six components, each normalized 0–100 across the cohort, each measuring one distinct way a person grows the network:
Engagement received (35%) — the heart of the score. Votes and replies your posts attract, where who engages matters: each interaction is multiplied by the engager's importance (0.5×–3.0×), replies count more than votes, and everything decays with time.
Onboarding (22%) — people you bring who become genuinely active. Not referral clicks: a referred account only earns you credit by actually posting within two weeks, surviving beyond a month, and attracting real engagement of their own. A hundred farmed signups that never post are worth almost exactly zero.
Centrality (13%) — your position in the graph above.
Consistency (12%) — showing up. Active weeks across the window, because a network is built by people who are still here on week ten.
Views (10%) — real impressions from the liketu apps, not a proxy. We instrumented actual viewability — a post has to be genuinely on screen (≥50% visible for ≥1 second) from an authenticated viewer to count, and viewers are value-weighted too. This signal is young and honest about it: its weight ramps in automatically as coverage grows. In the current window it's already built on 34,544 real impressions across 173 creators.
Engagement given (8%) — being a good citizen. Voting and replying to others counts, but deliberately can't carry you to the top alone.
The final score isn't quite a weighted sum — it's a weighted sum multiplied by integrity gates, and that distinction is where most of the design lives.
what you can't farm
Every scoring system is a magnet for gaming, and on Hive the attack surface is well known: vote rings, alt farms, stake-amplified self-dealing. LNV's answer is not one clever trick — it's that every input saturates or discounts exactly where gaming concentrates:
- A whale vote and an alt-stake vote saturate identically. Vote magnitude is log-scaled and capped at the cohort's 95th percentile — stake is used only for capping, never as a multiplier. Past the cap, more rshares buy nothing.
- Mutual back-scratching is discounted 60%. If you and I mostly engage each other, each interaction carries 0.4× the weight of one-directional, unprompted support.
- Concentrated support collapses. A diversity gate (Shannon entropy over who supports you) means engagement from two or three accounts — however intense — scores a fraction of the same volume from thirty independent people.
- Ring and sybil penalties multiply the whole score. Suspicious reciprocal clusters and connected components of coordinated accounts don't lose a few points in one column — their entire LNV is scaled down. You cannot max one clean input to escape a dirty graph.
- Farmed referrals score ~0 (quality-weighted onboarding), and impression farms score ~0 (distinct, authenticated, dwell-gated, value-weighted viewers).
Two things worth saying honestly. First, in practice the gates barely touch most people: the median support-diversity in the cohort is 0.93, and the ring gate meaningfully discounts about 3% of accounts. It's a fence at the cliff edge, not a tax on the meadow. Second, gates like these can graze genuine tight-knit friend groups — that's exactly the kind of calibration the validation phase exists for.
what the board looks like
The real spread of LNV across all 386 scored accounts in the July 1 run. Median 29.7; crossing 56 puts you in the top 10%.
Inside the app this lives in our Growth Console — a leaderboard we've been watching daily to validate the model before letting it touch anything user-facing:
The detail that matters most isn't visible in the ranking itself: every score ships with a receipt. For every account, every day, we store all six sub-scores, the raw signals underneath them, the diversity measure, and each penalty applied. There is no "the algorithm decided." Any rank on that board can be decomposed back into the exact human behavior that produced it — and that's also what keeps us honest when we tune the weights.
One receipt-level fact jumped out immediately and became my favorite thing about the current board: not a single account is earning onboarding credit right now. Twenty-two percent of the score — the second-biggest component — is sitting untouched, because almost nobody is bringing new people in a way the system can verify. The most valuable thing a liketu creator can do this month is also the least crowded: bring one real person who stays.
how it steers the vote
Here is where measurement becomes decision-making.
Your standing (percentile of LNV within the cohort) maps to a vote weight through three brackets — the bottom 60% earns 1–10%, the next 30% earns 11–20%, the top 10% earns 21–30% — interpolated smoothly inside each bracket so the vote tracks your score, not a cliff. On top of that curve sit the operational rules: a dynamic budget sizes votes to the day's supply of eligible posts (a quiet day boosts everyone up to 4×, a busy day trims so more creators get covered, and voting power is never left idle), posting five times a day decays each successive vote by half, moments are capped small, and a blacklist is absolute. The rollout is deliberately cautious — the engine runs as a suggestion layer first, sizing every vote with a human hand still on the switch, and flips to full auto only once we trust it.
The philosophical shift is the point: the vote follows demonstrated contribution to the network, not follower count, not stake, not who asked loudest in a comment section. If you want a bigger vote from us, the path is completely public: make things people genuinely engage with, show up consistently, engage beyond your circle, bring someone real.
And a version of it is already public-facing. The "people of liketu" rail on Discover — rising creators, welcomers, consistent posters — is LNV's first user-visible surface. Deliberately: names only, no scores. We celebrate the people; the number stays a tool.
the flywheel
Why go to all this trouble for a curation vote? Because of what the vote trains.
Every reward system teaches its community a strategy. Reward raw engagement and you teach engagement farming. Reward stake and you teach accumulation. Reward network value and the winning strategy becomes growing the network — because that's literally what the number measures:
- Rewards flow toward creators whose work pulls real attention from many independent directions → making genuinely engaging things becomes the best strategy.
- Onboarding is the biggest open lane → bringing real people becomes the best strategy.
- Consistency compounds → staying becomes the best strategy.
- Cross-community engagement beats circle-voting → the graph gets denser and healthier.
Each of those behaviors produces more users, more attention, more participation. More participation makes the graph richer, which makes the measurement sharper, which makes the votes better targeted, which strengthens the incentive — around and around. That's the positive feedback loop we're trying to spin up, and note what "quality" means inside it: not our taste in photography, but outcomes the network itself reports — people came, people stayed, people engaged of their own free will, from many directions. Quality as measured by the results we actually care about.
And none of this stops at our edge. liketu creators are Hive accounts; every onboarded person is a new Hive user; every vote, reply, and payout in our graph is an on-chain event; attention captured by liketu is attention captured by Hive. A curation system that reliably converts rewards into network growth is compounding value for the whole chain — we're just running the experiment on the community we know best. The model was deliberately designed to generalize beyond the liketu cohort, and whole-Hive scoring is the obvious next horizon.
Nine daily runs, five real accounts (anonymized). Scores are cohort-relative and recomputed from scratch every day — standing is earned continuously, never owned.
where this goes
LNV is young and we're treating it that way. The score recomputes daily and we watch it like a hawk. The weights are tunable and will be tuned. Views are still ramping toward their full weight as impression coverage grows. The importance-weighting is static in v1 (a full recursive PageRank-style version is built and switched off, waiting until we can interpret it responsibly). The vote engine keeps a human in the loop until it has earned autonomy. Every one of those caveats is stored in the open — in receipts, in degraded-run flags, in cohort-size labels — because a measurement system you can't audit is just vibes with extra steps.
But the direction is set. The interface should disappear behind the content; curation should disappear behind the community. What's left is a simple, public promise:
Grow the network and the network will notice. Bring someone real. Show up this week and the next. Make the thing people can't scroll past. The math is on your side now.
Explore liketu through our showcase pages
🔗 liketu.social — explore our features
🐾 liketu.social/wild — Wild, the field guide
🔍 liketu.social/shape — Shape, search by image
📲 liketu.social/download — get the android app
🏃 liketu.social/health — StepJar step pools
🗺️ liketu.social/atlas — Atlas
❓ liketu.social/faq — questions, answered
image-first. mobile-first. built on Hive.