Make the score answerable

Tags: scores, explainability, trust

A score does two jobs and is good at one. It ranks well. It explains nothing. So a user has exactly two available responses to it — believe it, or don't — and no third option where they check, disagree with the reasoning, and carry on using the product.

Why the first bad call is terminal

Watch the trust curve on an opaque score and it has a distinctive shape. Early on people extend credit, because a number looks like it was computed and computed things are usually right. That credit survives a while. Then the product recommends something obviously wrong for them.

At that moment a user with an answerable score does something ordinary: they open it, see that it weighted a thing they had rated highly by accident, correct it or discount it, and move on with a more accurate model of what the score is good for. Trust goes down slightly and becomes better calibrated.

A user with an opaque score can do none of that. There is nothing to open, so the only available update is about the score as a whole — and it goes to zero, permanently, on one sample. The score never recovers because there was never any mechanism by which it could.

Answerability is not a transparency nicety. It is the repair mechanism. Without it, every score is one bad recommendation away from being ignored forever, and you will never know which recommendation it was.

What answerable means, concretely

Asking why this number should return something the user can check against what they already know, and ideally argue with.

In their terms, not the model's. "Because you rated three shows with the same lead" is checkable against their own memory. A feature weight is not, however faithful it is to the computation.
Specific enough to be wrong. An explanation nobody could disagree with is decoration. "Based on your viewing history" is that. "Because you finished all four seasons of this and rated it five stars" can be met with I watched that with my sister and hated it, which is the whole point.
Connected to a control, where possible. The best version lets the disagreement do something. If the reason is wrong, the user should be able to say so and see the score move — which is also how you find out your factors are wrong.

Confidence is not the answer to this

This is where the published catalogs and I part company, so it is worth stating plainly.

A confidence percentage is a number about a number. The user cannot check it — there is nothing to check it against — and it does not tell them what to do differently. Ninety-three percent confident and sixty percent confident produce the same available action, which is to take it or leave it. It has the shape of transparency, and it discharges the feeling of having explained, while leaving the user exactly where they were.

It is also, quietly, a second opaque score layered on the first. If the problem is that people cannot interrogate the model's judgment, adding the model's judgment about its judgment does not help.

The existing mark what the product vouches for makes the same argument from the other end, about prose claims: confidence scores are the wrong instrument, and the right question is what the interface will stand behind.

Or don't show it

The other half of the fork, and the cheaper one.

A score used purely for ranking does not need to be answerable — it needs to be invisible. If the model orders a list well, ship the ordered list. The number is doing its job silently, and nobody is owed an explanation for an ordering they did not know was computed.

Surfacing the number as a judgment is what creates the obligation. So the decision is not "should this be explainable" but "is this a judgment I am prepared to defend, or a sort key I let leak into the UI". Most product scores are the second thing wearing the clothes of the first.

Grounded in

WatchWorthy turned a decade of Ranker's crowd votes — over a billion of them, five thousand lists about what people watch — into a Worthy Score. Crucially it shipped alongside a visible taste profile, so a recommendation came with the shape of why it fit rather than only the assertion that it did.

That combination carried a consumer product for six years. It shipped in March 2020 and is still shipping: 4.6 stars across roughly four thousand ratings, now spanning fifty-plus services. A score people could partly see the reasoning behind survived contact with a very large number of individual bad recommendations, which is the only real test this pattern has.

The rebuild exists because of the remaining gap, and it is the honest version of the lesson. The 2019 app could rank. It could not be asked. A visible taste profile tells you the general shape of the model; it does not answer why this one, and the difference between those two is where the pattern actually lives.

Anti-patterns

A confidence percentage as the explanation. Covered above, and the most common version of getting this wrong while believing you got it right.
Precision as authority. A 93.7% match asserts a resolution the underlying judgment does not have. See draw the certainty boundary — significant figures are read as a claim.
Explanations in model terms. Feature weights and attention maps are faithful and useless. The user is not debugging your model.
An explanation with nothing to disagree with. "Based on your activity" is a sentence, not a reason.
A score that changes with no account of why. If it moved from 88 to 61 overnight, that is the moment answerability was worth having.
Explaining after the fact with a different model. A second system generating plausible-sounding reasons for the first system's output is not an explanation. It is a fluent guess about a number, and it will be wrong in ways nobody can detect.

The smallest version worth building

For each score you show, write down the three factors that most often move it. Show the ones that actually applied to this item, in the words the user would use.

That is not a full attribution and does not need to be. It converts the number from an assertion into a claim with a surface — and a claim with a surface can be corrected, which is the only way trust in it ever recovers.

Mark what the product vouches for — the prose version of the same decision, and where the argument against confidence scores was first made here.
Draw the certainty boundary — how much precision a number is entitled to display.
Evaluation loops — how you find out whether the factors you are showing are the ones that actually moved it.