The Floor You Cannot Buy Back

129 out of 140. Ninety-two percent. The review said no.

The surface was a dashboard for volunteers who run a broadcast platform. They open it rarely and decide fast. Is anything waiting on me? Is the spend on track? By the third review the build answered both.

It had not started there. The first review scored it 79 and called for a redesign. Six blockers. One chart that drew the quarterly budget as spent when nothing had been spent. Accessibility scored 1 out of 4. The second review said 110, two points under the ship line. The third said 129. The fourth said 133 and let it go.

The third one is the one I keep thinking about.

A weighted score is a market

The rubric has nine dimensions, each scored 0 to 4, each carrying a weight. Accessibility carries 4 of the 35 weight points. That is about 11 percent of the total.

So the arithmetic is simple, and a little ugly. At the 129 review, accessibility sat at 2. Drop it to 1 and the total becomes 125. That is 89 percent, comfortably over the ship line the reviewer applied, 112. Eight other dimensions were at 3 or 4, and they paid for it.

This is how inaccessible products ship. Nobody decides to ship one. The total looks healthy, the strong dimensions carry the weak one, and the weak one never gets a vote.

A weighted score is a market. Every point is fungible. Spend enough on hierarchy and you can buy your way past a failure in anything else.

It was off by 0.07

The failure at 129 was small. An accessibility scanner measured two text nodes at 4.43 to 1 against a floor of 4.5. One was the muted footer line of the budget widget. The other was the description in an empty state. A separate check found an unnamed, focusable wrapper around the refresh button.

The builder was an AI coding agent. It had already triaged the contrast pair out. The muted color was a shared token, it argued, used the same way on older widgets, so not this surface's problem.

The reviewer agreed on who should fix it and disagreed on whether it counted. A contrast gate measures rendered values. It does not care which stylesheet owns the color. I understood the builder's argument. It is a reasonable argument for a total. It is the wrong one for a floor.

The fix took less than a point of lightness. The shared token moved from 55.17 to 54.5 percent, and the two nodes landed at 4.53 and 4.58. The wrapper came off the button, which already carried its own label.

Four reviews, two days. One of them was spent on a 0.07 gap.

Four dimensions are not for sale

The tool I built, Considered, does not let this happen. It scores the same nine dimensions. Then it floor-gates some of them. The reviewer in this story applied four: purpose fidelity, information architecture, visual hierarchy, and accessibility. A surface ships only if each scores at least 3, whatever the total says. The tool's policy file has since been recalibrated. The ship line is now 120, and craft joined the floors as a fifth.

The rubric spells out the reason for the first one. A beautiful, accessible dashboard that answers the wrong question is worth less than an ugly one that answers the right one. Strength elsewhere cannot rescue it.

I had a 92 percent and a tool that refused it. My first reaction was to look for the exception. That reaction is the argument for a floor, because the exception always looks reasonable once the number is yours.

Weights say how much you value things. Floors say what you will not trade. Most design organizations I have seen have weights and call them standards.

A floor only works if somebody measures it

There is a second half. At the 110 review the accessibility score was capped at 3, because nothing had been measured. Seven of twelve accessibility gates were blocked, with no scanner run, no keyboard pass, no screen-reader check. An unverified gate is not a passing one, so the score could not go higher.

Then the evidence arrived. The measuring turned two unknowns into two known failures, and the score fell from a generous 3 to a true 2. The reviewer called that the honest outcome of measuring. That looks like regress. It was the first time the number was real.

The last review gave 133, not 140. Accessibility rose to 3 and stopped, because two gates were still unmeasured. The review said so and shipped anyway, because the floor was met and what remained cost more than it returned.

A floor with nobody measuring is a wish. A floor with a loud failure is a policy.

Decide before the review

Every team has a version of this. A launch checklist that quietly becomes advisory. A security review that gets waived when the date is tight. Each began as a floor and was converted into a weight, one reasonable exception at a time.

Pick the few things you would rather miss a date over than trade, and make them floors, not weights. Kill Criteria Up Front argues for deciding early. A floor is what deciding early looks like inside a rubric.

The dashboard shipped at 133. It would have shipped sooner at 129, and been worse.