The Feature I Will Not Ship

The most requested feature in the thing I am building is a single green checkmark. One answer: are you fine, yes or no.

It is also the feature I am most confident we should never ship, and working out why took me longer than it should have.

The request is completely reasonable

I want to start here, because it would be easy to frame this as users wanting the wrong thing. They do not.

An operator has a document, a shelf full of product, limited time, and no appetite for reading regulatory text. They want to know whether they have a problem. A list of findings is work. A green checkmark is an answer.

If I could give them one honestly, I would.

Why I cannot

Producing a single answer requires collapsing things I genuinely do not know into a confident output.

Here is the concrete case that settled it for me. One threshold in our domain is verified as text — we can cite it precisely against the adopted rules, dual-sourced, with a date. Its enforcement posture is a separate question, and honestly unsettled.

Now: what does the checkmark say?

If it goes green, I have asserted that an open question is closed. If it goes red, I have asserted a violation the record does not support. There is no third colour that means "this rule is real and how it is being enforced is genuinely unresolved" — because the entire premise of the checkmark is that a third state does not exist.

So our checker reports findings per requirement rather than overall. It never states that a product may be offered for sale. Where a rule's enforcement posture is pending, it returns NEEDS_REVIEW.

The uncomfortable part

That is a worse demo. Measurably.

Users occasionally tell me it is a worse product, and I do not think they are being unreasonable — from where they sit, they asked a question and got a list.

The argument that holds it in place is about who carries the consequence. If I render a verdict and I am wrong, I have a difficult conversation. The operator has an enforcement action, a recall, or a business that stops. The asymmetry is total, and it runs entirely against the person with the least information.

A tool that tells you which specific requirement is unmet is doing work you can act on. A tool that tells you that you are "compliant" is making a promise it has no standing to make, to the one party who cannot check it.

What I think this generalises to

Any system that summarises risk for someone else faces the same choice, and most resolve it the other way, because the summarised version is what sells.

The test I have landed on is simple: would I be comfortable if the user could see exactly how confident the system actually was?

If the honest confidence would embarrass the interface, the interface is lying — and the lie is being told to whoever is holding the consequence.