How far from the obvious

One scale

Zero is the typical answer: the centre of many independent plain answers to the same question. One unit is the spread of a realistic pool of ideas for that question, so about 95% of a pool falls within two units. We call the distance W. Every measurement so far is unsigned. It says how far an idea sits from the typical answer, not which way it leans, and a signed version is still to come.

On one brief, the scale spans 130 model answers to the eVolo Skyscraper Competition, from the pile of tidal towers at zero to a tower whose every structural joint is a room, at 2.09.

A dial that does what it says

Asked for five ideas at once at a given level of strangeness, a model drifts: later ideas in the list get weirder regardless of what was asked. Asked for one idea per call, the requested level and an independent rating of how unexpected the idea is agree at a Spearman correlation of 0.89. We still measure every idea after it is written rather than trusting the request.

Where people land

Where people land on the eVolo brief

0 the typical answer0.66 the founder’s picks peaked1.43 median of what the jury awarded2.09 farthest model answer

00.511.52

distance from the typical answer

Fourteen picks give a wide range. Even so, the founder and the jury sit far apart on the same brief.

A lean, not a sweet spot

A much larger test since then changes how we read the founder’s peak. Among 45,136 people who each rated at least 40 of the same 100 jokes, each person had a small, stable lean toward or away from the unusual once joke length was taken out (split-half correlation 0.11 to 0.16), and people split evenly between the two directions. What we did not find is a favourite distance: no one’s ratings peaked in the middle more reliably than chance, and synthetic people who only lean reproduce everything we saw. So, within ordinary ideas, we treat a person’s distance as a direction to rank by, not a target to aim at. Whether far-out ideas behave differently is still untested.

Juries differ by medium

We scored three large text contests whose full entry lists and outcomes are public. Short-story juries on Reedsy, across 119,638 stories in 365 contests, recognised the most atypical openings 2.3 times as often as the most typical tenth. The crowd rating New Yorker cartoon captions, across 2.18 million captions, favoured the typical ones, and that holds once caption length is controlled; the Reedsy result has not yet been checked for length. ICLR reviewers, across 13,585 papers, showed almost no relation. The research literature points the same way: work with a conventional core and one unusual element tends to do best, in science papers, in pop songs and, in our own analysis of 15,539 artists on Wikidata, in artistic fame.

What it means

There is no single right amount of weirdness. How much unusualness wins depends on the medium and the judge, and each person leans their own way. So the lean has to be learned per person, from that person’s picks, and the selector matters most.

Next: One idea in thirty outfits