This page exists so that everything else here can be checked. It was written before we had enough traffic to publish a single first-party statistic, which is the only honest moment to write it: rules set after seeing the data are not rules, they are decoration.
What is set out here, and what is not published anywhere
This page is about measurement: what counts as an error, which denominator applies to which rate, how large a sample has to be before it means anything, and what we refuse to say at any sample size. All of that is published, because a claim you cannot check is not worth making, and a threshold announced afterwards is not a threshold.
How JunkGate arrives at a verdict is not published, here or anywhere else. Saying so plainly is better than implying there is nothing to say: this is a field where the people sending the material read the vendors' websites, and a public account of what a system looks at is a public account of what to send instead. The businesses using JunkGate would pay for that, not us.
We publish how our claims are measured, not the signals used to reach a verdict. Where those two pull against each other, we take the side that protects the enquiries.
The two mistakes, and why we never average them
A filter makes two kinds of error, and they are not comparable.
- A real enquiry wrongly held
- Costs the business a customer, and usually invisibly. Nobody knows the message existed, so nobody knows it was lost. This is the fatal error.
- Junk that got through
- Costs the business a few seconds of irritation. This is the annoying error.
An accuracy percentage adds these together, which means it hides the one that costs money behind the one that does not. A filter that holds one real enquiry in fifty and catches everything else can advertise 98% accuracy while quietly losing a business its best lead of the month.
We do not publish a single accuracy figure for JunkGate, and we will not. The two rates are reported separately, each with its own denominator, or they are not reported at all.
What has to be true before we publish a number from our own system
A claim about what we see in contact-form traffic is only worth reading if the sample behind it is stated. Ours are pre-committed and public:
- At least 100 eligible submissions behind any pattern claim.
- Spanning at least 30 days. A large sample gathered over one afternoon describes an afternoon, not a pattern.
- Every proportion carries its denominator in the same sentence.
- Exclusions are counted and stated: our own tests, duplicates, and anything generated by us rather than sent to a customer.
These thresholds are enforced by the software that builds this site, not by our good intentions. A document that claims a first-party number without declaring its sample size and period does not render with a warning; it stops the site from starting. We would rather take the site down than publish a number we cannot stand behind.
They are also not adjusted downward to make a thin dataset publishable. If we ever change them, the change and its date will appear on this page, and the reason will be stated.
What we will never publish
- The contents of anyone's messages. The people who fill in our customers' contact forms never agreed to anything with us. Their words are not our material. Where we quote a scam, it is one sent to an address of ours, or it is reconstructed and labelled as such.
- Anything identifying a business or a sender, including through a combination of details that would narrow it to one.
- Cell counts small enough to identify. Where a breakdown would rest on a handful of messages, we suppress it rather than round it.
- Someone else's statistic presented as ours. External figures appear in a visually distinct block, attributed, with the date we read the source.
Corrections
Pages keep their addresses. When we revise one materially, the revision is dated and the change is described at the foot of the page. When we get something wrong, we correct it in place, say so, and add it to the corrections log with what we said, what was actually true, and the date it changed. We do not delete a page to make an error disappear, because a reference that quietly rewrites itself is not a reference.
A revision is not a correction. Rewriting a page is ordinary; publishing something that was not true is not, and only the second kind reaches the log.
Where our own knowledge comes from
We keep a working picture of how the people sending this material are operating now. It is dated, refreshed on a slow cadence as tradecraft changes, and published here as the threat brief. Publishing it costs us almost nothing, and a reader can see how old it is, which is the part that matters.
Revisions
10 August 2026. Passages describing how our own system reaches a decision were removed, and the boundary that replaced them is now stated in full at the top of this page. The definitions, the sample thresholds and the publication refusals are unchanged. A corrections log was added and is linked above.