Evidence
Methods and evidence
probe, scan, cert, interpolated, feed: what each method establishes and what it does not.
Every claim carries the method that produced it. The methods are not interchangeable; the ordering below reflects evidential strength.
| Method | What happened | Strength |
|---|---|---|
probe | Traffic was routed through the provider and egressed here | Direct observation |
cert | The TLS certificate issuer names the provider | Strong. Operators sign with their own CA |
scan | The host responded on a port used by this infrastructure | Presence established, egress not proven |
interpolated | Not observed itself; other addresses of its announced prefix were, and the prefix’s providers are reported | Inference. Confidence from the share of the prefix observed |
feed | An external list asserts it | Third-party measurement. Publisher and retrieval date attached |
inferred | Derived from neighbouring observations | Weakest. Treat as a lead |
Interpolated addresses
An address we never observed is answered from the announced prefix it sits in — the most specific one the routing table carries — whenever any of our sources knows at least one VPN exit in that prefix. The prefix’s figures are the union of all sources as computed after the last sync — the same the prefix search shows: each known address once, by the provider of its most recent sighting. Providers and weights follow from those counts; network names the prefix; evidence.coverage is the share known, computed by us: distinct known addresses divided by the prefix’s size. First and last seen are the prefix’s first and last sighting in any source.
| Share of the prefix observed | Confidence |
|---|---|
| More than 50 % | High (3) |
| 10 % to 50 % | Medium (2) |
| 1 % to 10 % | Low (1) |
| Less than 1 % | Very low (0) |
A very low answer — under 1 % of the prefix known as VPN exits — is not anonymous: it says a provider was seen in the network, not that this address is one, and the address page says “Not observed”. Nothing is filtered out by confidence — set your own threshold on the single and the bulk lookup: min_confidence=1 leaves the very low answers out, min_confidence=2 the low ones too. Where the prefix is shared between providers, the answer is shared too; read the weights, not only the leader.
A provider's network list as a blocklist. The list names every announced prefix in which the provider was seen, however large and however little of it the provider uses: a cloud's /11 holding one of its servers is on it. Its confidence counts the provider's addresses there — whether the provider is present —, not their share of the block. Before blocking whole blocks, read coverage: the provider's addresses over the block's size, in the same bands as interpolation. min_coverage=2 keeps the blocks of which at least 10 % is the provider's; max_size and exclude_datacenter leave out large and cloud blocks. Without filters the list stays complete.
What the confidence of an interpolated answer says. It states how much of the prefix we observed as VPN exits — the share, with its numbers, is in confidenceBasis — and nothing more. It is not a measured hit rate: we do not claim that, say, nine of ten medium answers are right. Two prefixes with the same share can differ. One is rented whole by a provider that uses only part of it at a time, so the addresses we have not seen are its too; the other is split by its operator among several customers, so the rest belongs to others. The share cannot tell them apart. Choose your threshold with that in mind: where a false alarm is costly, act on high confidence or on observed addresses only.
Confidence
Every answer states its confidence — high (3), medium (2), low (1) or very low (0) — and how it was computed, in evidence.confidenceBasis and on the address page. The rule depends on what the answer rests on:
| Rests on | High (3) | Medium (2) | Low (1) | Very low (0) |
|---|---|---|---|---|
| Our probe events for the address | 50 or more | 5 to 49 | fewer than 5 | — |
| A certificate naming the provider | always | — | — | — |
| The enriched snapshot alone | — | always | — | — |
| Interpolation from the prefix | more than 50 % observed | 10–50 % | 1–10 % | less than 1 % |
| External lists alone | — | — | always | — |
Feeds never become measurements
External lists supplement gaps and never override a measurement. Attribution weights derive from our own probing and scanning only. Where a list names a provider, it appears under method: "feed" with the publisher, keeping retrieved data distinguishable from measured data. See External sources.