How ICPulse Measures Governance Influence
What this page explains
ICPulse publishes influence estimates for the NNS's registered known neurons — how much voting power actually moves when each of them casts a vote. This page explains where that number comes from, why it can be trusted, and what its limits are. We believe a metric you can't interrogate is a metric you shouldn't trust.
The problem: follower counts are not public
On the Internet Computer's Network Nervous System, any neuron can follow another neuron: when the followee votes, the follower's voting power is cast automatically. Following relationships, however, are private by design. They are stored on the follower's side, and no API, dashboard, or canister exposes "how many neurons follow neuron X" or "how much voting power stands behind it."
So the single most important fact about governance power on the NNS — who actually moves how much — is not directly observable. It can only be inferred.
Our approach: observed cause, observed effect
What is public, verifiable, and updated continuously on-chain:
- Proposal tallies — the running total of Yes/No voting power on every proposal.
- Known neurons' ballots — the governance canister publicly records how each registered known neuron voted on each proposal, precisely so that voters can evaluate whom to follow.
ICPulse runs monitoring infrastructure that continuously watches both. When a known neuron casts a vote, the NNS automatically cascades its followers' votes into the tally. By comparing the tally immediately before and shortly after that single event, we observe the size of the wave that one neuron set in motion: its own voting power plus its followers'. Subtracting the neuron's own (public) voting power leaves an estimate of its follower power.
Why the estimate is credible
Quiet-window discipline. A tally jump only identifies a cause cleanly if nothing else significant happened in the same time window. Most voting power on a typical proposal arrives in one early wave (triggered by the largest default-followed neurons). We deliberately measure in the quiet periods outside such waves, where late voters arrive one at a time.
Aggressive disqualification. Every measurement window is checked against the full public record of known-neuron ballots. If any other known neuron voted near the same window, the measurement is discarded — no exceptions, no weighting tricks. We prefer fewer, cleaner data points over more, noisier ones.
Open-voting exclusion. Votes cast while a proposal is still open for voting are never counted as measurements, even in an otherwise quiet window. During the open period, the network's background voting — thousands of regular neurons voting directly or through private followings — moves the tally continuously, and a delta measured then reflects that ambient wave rather than one neuron's followers. We verified this empirically: simultaneous votes by one known neuron on three different open proposals produced nearly identical deltas — the signature of background noise, not of three different follower sets. Ballots cast during the main decision wave — when many neurons vote in overlapping windows — are discarded as contaminated. Only votes that land in quiet windows, with no other known-neuron activity nearby and low background voting, qualify as clean measurements.
Distributions, not anecdotes. No single measurement is ever published as the answer. Each figure you see is the median of many independent clean measurements, and we display the sample size (n) beside it. Small-n figures are labeled as provisional.
Per-topic resolution. NNS following is configured per governance topic. A neuron may command a large following on protocol-management topics and a small one on Governance motions — or vice versa. ICPulse therefore measures and reports influence per topic, never as one blended number.
Cross-validation against the official dashboard. Every data source we consume is an official, public Internet Computer API — the same infrastructure that powers the official dashboard at dashboard.internetcomputer.org. Key figures are cross-checked against the official IC dashboard where a comparable figure exists; not every widget has a formal reconciliation step. Each widget's footer names its sources.
What this metric is — and is not
- It is an estimate inferred from public on-chain effects, not a ledger entry. There is no ground truth to compare against, because the ground truth is private by protocol design.
- It is a snapshot in time. Followings change, voting power decays for neurons that don't reconfirm participation, and estimates drift accordingly. Every figure carries its measurement dates.
- It measures realized influence (power that actually moved when the neuron voted), which is the economically meaningful quantity — not the number of follower neurons, which conflates many tiny neurons with a few large ones.
- Neurons that habitually vote early, inside the main decision wave, are harder to measure cleanly; their sample sizes will be smaller and their estimates wider. We show this rather than hide it.
What we don't publish
The exact measurement windows, disqualification thresholds, sampling cadence, and reconciliation internals are ICPulse's engineering. Publishing them wouldn't help you evaluate the numbers — the validation logic above does that — and the exact heuristics evolve and are intentionally not published in full detail. What we commit to: every published figure is reproducible in principle from public on-chain data, and our methodology's logic is fully described on this page.
Exchange address attribution
A small set of ICP addresses is labeled as exchange-owned on /transfers and in the Whale Feed. Each label carries a verification date and one of two statuses: verified or inferred. We describe the method here; the specific wallets and exchange names are not listed on this page.
Verified means the address was confirmed via controlled transfers we executed and observed on-chain. Inferred means the address was attributed via transaction-pattern analysis — zero-balance instant pass-through relays, funding chains, and similar structural signatures — anchored to addresses that were themselves verified. Inferred labels are best-effort and carry less certainty than verified ones.
Labels can become stale. Exchanges rotate wallets, retire old ones, and route through fresh intermediaries, so a label that was accurate on its verification date may no longer describe the current occupant of that address. The verification date is shown alongside each label so you can judge how recently it was confirmed.
Questions or challenges
If you believe a published figure is wrong, we want to know. Every widget lists its data sources; challenge us with on-chain evidence and we'll investigate publicly.