I design typefaces. That is why this framework does not trust your survey.

Where the idea actually came from

Type design trains you to care about things nobody notices. A terminal cut two degrees short. A sidebearing one unit too tight. Nobody reading a paragraph consciously registers any of it, and every one of them feels it in how long they stay on the page.

You do not fix that by asking readers what they think of the sidebearings. They have no idea. You fix it by measuring the thing itself.

Then I spent a decade in product rooms doing the opposite. A satisfaction score going up, a funnel going down, and a table of intelligent people quietly selecting whichever number supported the roadmap they had already agreed on.

Silent friction is the typographic problem wearing different clothes. A cost the user pays every session and never registers as cost. Ask them and they will honestly tell you it is fine, because to them it is. That is the thing RUCF was built to find, and it is why the framework refuses to average behaviour and perception into one number. Averaging them destroys the only signal worth having.

What the work taught me

Ten years of client work across fintech, gaming, publishing and institutional communication, including a stretch as creative director owning brand and product design end to end. Some of it was reporting and packaging for organisations whose entire job is auditing their own claims.

That work teaches you how fast a number turns into a decoration. Nobody falsifies anything. The figure is real, the chart is accurate, and by the third revision it has stopped being evidence and started being a design element that supports a sentence somebody wrote first.

RUCF is built against that. It makes you declare how you know each figure, and it weights expert judgment at 0.30, mine included. It is deliberately unflattering. A framework that agrees with you is a marketing asset, not an instrument.

On AI, and why it is capped

I studied Designing and Building AI Products and Services at MIT, largely to work out where classification can be trusted and where it cannot.

The short answer is that it is very good at volume and unreliable at nuance, and the failure mode is confident rather than obvious. So anything a model classified enters this framework at 0.50 confidence and stays there until a human has validated a sample. Every other framework treats AI as a feature to advertise. Here it is a discount applied in the arithmetic.

Stress testing

I mentor designers on ADPList, which is where a lot of this language got tested on people with no reason to be polite about it. Several of the quadrant names and most of the plain-English criteria exist because someone in a session told me the previous version made no sense.

About this site

The typography is doing real work. Scores are set in a monospaced face with tabular figures so a number never jitters as it counts up. No score appears anywhere without its confidence attached, including in the marketing copy. The confidence bar is the most repeated mark on the site by a wide margin.

If you remember one graphic from RUCF, it should be four segments with two of them empty.

The mark in question.

The full argument

The sourced case for the framework — attitude–behaviour research, the evaluator effect, evidence grading, tail reporting, and the conditions under which the whole thing should be rejected — is written up as the manifesto.

Elsewhere

Work, typefaces and older projects live at ipassas.com.