3 ms·
A long time ago, in a faraway, and mythical country, which we'll call China, everyone wanted to know how long the Emperor's nose was. Of course to be seen even
by dfranke 17y ago
A long time ago, in a faraway, and mythical country, which we'll call China, everyone wanted to know how long the Emperor's nose was. Of course to be seen even trying to look closely at the Emperor's visage - let alone to hold up a (different!) ruler to it - would have invited instant, or I should say, far from instant, death. But so many people were curious, that a group of sages got together to look for a method of finding the answer, and this is what they came up with.
Questionnaires were printed and sent out in bundles to cooperating village chiefs, who distributed them to the peasants. Literacy was at a sufficient level that most were able to complete the single question, which was, of course: "How long do you think the Emperor's nose is?"
When the forms were collected, mathematicians added up all the values, and divided by the number of forms. Thus it was known that the length of the Emperor's nose was 6.734602 cm. The complete set of data was of course preserved, and many years later, with advances in statistical understanding more advanced mathematicians pointed out that fringe values - obviously the product of deranged minds - were distorting the honest opinions of the rest, and by eliminating them and using the very latest numerical modelling techniques, the mathematicians corrected this value to 4.980403 cm. To this day, no-one has produced a better estimate.
- hs 17y agoyour china emperor story only suggests that the parent post's averaging scoring scheme consist of reviewers who dream about a 'test-taker' and then give a score (repeat as many as test takers there is) ... if they have access to takers' answers and keys, then averaging could be done much more objectively the china emperor: that's maybe because there's no photograph of the emperor circulating. if there is, one can make comparison to eyes, age, coin, ground-legal-measure etc and clean out the outlier data. averaging the rest can give good enough estimate but not super accurate.