Fable
The difficulty for a copyist to guess the next character
Songfeng Hall had a group of copyists who specialized in filling missing characters in old scrolls. In the early years, there were few gaps and the surrounding sentences were clear, so the old masters could mostly guess which character should be filled in by looking at the context. Later, the hall collected foreign poems, prescriptions, account books, and ship diagrams. There were more and more missing characters, and the styles were mixed. The copyists often stopped, because the same position might have several reasonable answers.
The curator's naive fix was to make everyone copy faster. Whoever paused least was rated the best copyist. As a result, many began forcing in common characters; the exam papers read smoothly, but prescriptions and account books were corrupted.
A textual critic changed his approach. He not only looked at what characters were ultimately supplied, but also at how hesitant the copyist was before each gap: if the context was clear enough, good answers should be very concentrated; if the possibilities were scattered, it meant this volume was hard to predict.
He gave each copyist a batch of standard damaged scrolls, had them predict character by character, and recorded their confidence in the correct character. It was not about celebrating one lucky guess, but about seeing how hard the whole scroll was on average for them to predict.
One copyist performed well on poetry but always hesitated on prescriptions. Another copyist could reliably predict ledger terminology but could not write beautiful prose. The curator began to understand that low hesitation does not mean good at all tasks.
Later someone wanted to use this score to decide who could independently repair books. The collator stopped them: it can measure whether language prediction fits the text, but it cannot guarantee factual correctness, correct tool use, or task completion.
So Songfeng Hall put this score in its proper place: use it when training basic language skills, use it when comparing similar texts; before actual delivery, human collation, fact-checking and purpose acceptance are still required. The collator also noticed that a copyist hesitated less on familiar styles and more when switching to unfamiliar volumes. The score could therefore reveal whether a text was close to his training experience, but it could not understand drug properties or history for him. The collator also reminded the curator that low hesitation is sometimes dangerous. If the copyist was especially confident about a wrong answer, the score would reveal that he did not put the correct character first; if the whole volume's style was unfamiliar, increased hesitation might simply mean the material was beyond his familiar range.
The curator finally understood that knowing how predictable a sentence is is useful, but it is only a ruler for understanding the flow of text, not a scale for measuring all intelligence.