Almost everything in this category treats its two failure modes as roughly equivalent and optimises the sum. That is a modelling error with a real cost, and correcting it changes what a business should be willing to accept, not just how a product is tuned.
The asymmetry
Consider what each mistake actually costs the business.
Junk that gets through costs a few seconds. The owner sees it, recognises it, deletes it, and is mildly annoyed. The cost is real but small, immediate, and visible.
A real enquiry wrongly held costs a customer. Someone wanted to buy something and was silently prevented from making contact. They conclude the business ignored them and go elsewhere.
The second is worse in magnitude, and worse in a way that compounds: it is invisible. The owner does not know the message existed. There is no notification of the absence of a notification. A filter can be quietly costing a business its best lead every month and produce no evidence of it at all, while its dashboard shows an impressive count of things stopped.
Why a single accuracy figure hides exactly the wrong thing
An accuracy percentage adds the two together. Since junk vastly outnumbers genuine enquiries in most contact-form traffic, that total is dominated by the cheap error, and the expensive one can move a great deal without troubling the headline number.
A filter holding one real enquiry in fifty, while catching nearly all junk, will advertise a number in the high nineties. That number is arithmetically true and practically misleading, because the fraction it is hiding is the fraction that costs money.
This is why the two rates belong apart, each with its own denominator, and why we do not publish a combined figure for our own product. It is a genuine cost: one impressive number markets better than two honest ones. A filter that cannot be audited on the error that matters should not be trusted with a business's customers, including ours.
What a business is entitled to demand
The asymmetry is not an opinion about tuning. It settles several questions outright, and each of them is something an owner can insist on without knowing anything technical. These are worth putting to anyone selling you this, us included.
Unsure has to mean delivered
Not every message is clear-cut, and the ones that are not are exactly where a wrong answer is most expensive. The only defensible resolution is that an unclear message reaches the business. Anything that forces a call in both directions is choosing to guess on the cases where guessing costs the most.
Broken has to mean delivered too
An outage, a timeout, a failure of any kind on the vendor's side: every one of them has to end with the message arriving. A filter that fails closed converts its own downtime into the customer's lost sale, which is an outrageous thing to do to somebody's business. That includes billing. Nothing should ever be held because a limit was reached or an invoice was late.
Held must never mean deleted
A held message stays visible and releasable, and releasing it has to deliver the original notification unchanged rather than a reconstruction. If the recovery path is lossy, the promise is not really a promise.
It has to earn the right to act
Nothing should begin holding messages on the day it is installed, on a site it has never seen. Watching first and holding nothing costs a few days and converts an act of faith into an observation. Switching it on afterwards is the owner's decision to make.
The owner has to be able to overrule it
Everything held stays visible to the business it was addressed to, and any of it can be put back in one action, without asking anybody. A verdict nobody can reverse is a verdict nobody can correct, and a filter that cannot be corrected is eventually switched off entirely, which helps nobody.
The objection, and it is a fair one
Erring this hard against one error class lets more junk through than a more aggressive filter would. That is true, and it is the trade we have deliberately made.
The reason is that the two errors are not merely different in size, they are different in recoverability. Junk in an inbox is a visible problem with an obvious fix: the owner sees it, says so, and it can be put right. A wrongly held enquiry produces no signal at all. Given a choice between erring toward a mistake that reports itself and one that does not, the first is the only defensible direction.
The right way to make this argument is not to claim it has been solved. It is to state the trade, keep the two numbers apart, and let the person whose customers are at stake judge it.
Revisions
10 August 2026. Substantially rewritten. This page previously argued the asymmetry by describing how our own system is built, which we no longer publish. The argument is unchanged and the conclusions are unchanged; they are now put as what a business should demand rather than as an account of our internals. We publish how our claims are measured, not the signals used to reach a verdict.