v6 did not make the AI smarter. Across three blind arms, five runs each, v5 and v6 scored the same on catching problems.
What changed is whether the report has anywhere to put what the reader felt.
A reader finishes your manuscript, and everything they felt has to fit into a form with no boxes for most of it. Whatever does not fit disappears. v6 adds the boxes.
Manuscripts that are not in Chinese finally get the right language
This is the one to lead with.
Under v5, a manuscript written in English came back as a report that was 83% to 85% Chinese. The model speaks English perfectly well. The problem was on our side: our instructions unconditionally included a block of Chinese, and the readers did as they were told.
Measured on v6: an English manuscript now produces a report with 0% Chinese. Japanese, Korean and Spanish were each tested separately and all came back in the language they should be in. A Spanish manuscript used to come back sprinkled with Chinese characters. Now there are none.
“Surprised” and “absorbed” can be said out loud for the first time
v5’s emotion list had no surprised, no moved, no satisfied. The attention field had exactly one positive option: engaged.
So when a reader was genuinely caught by one of your paragraphs, the report could only say engaged. That was the only word available to it.
v6 adds absorbed, a level above engaged, closer to “could not put it down”, plus three positive emotions. What the readers feel has not changed. What changed is that the report can now write it down.
Foreshadowing gets its own analysis, in six kinds
Foreshadowing used to be buried inside the paragraph-by-paragraph reactions. No field answered the question you actually care about: was this set-up any good?
Every piece of foreshadowing now gets a verdict:
| Verdict | What it means |
|---|---|
| Well executed | Surprising but inevitable, and fair to the reader |
| Never fired | Heavily set up, then never mentioned again. The reader invested for nothing |
| Telegraphed | The reader guessed it long ago, so the reveal had no charge left |
| Unearned | Cashed in with no set-up. The reader feels cheated |
| Payoff unmarked | Set up and paid off, but too fast. The reader understood it and missed feeling it |
| Deliberate hook | Left open on purpose for the next book. Not a flaw |
The fifth one is worth a note. Writers have told us “the surprise of the reveal never showed up in the report”. That thing had no name until now.
Honestly, though: this is the verdict the model is still worst at. Most of the time it still calls those well executed. The field and the definition are in place; the judgement will have to wait for a stronger model.
Praise finally has to be specific
Problems had 40 slots and 9 structured fields. Strengths had one sentence.
The result was predictable. The tool was precise about what was wrong and vague about what was working.
From v6, praise has to name the passage, quote the line, and say why it works. Praise that cannot produce the quote does not count.
Three things fixed along the way
Two readers had been running with a broken instruction. Two of the nine were being asked to fill in a value that was not even valid. The retry mechanism kept teaching them back to a correct answer, so nothing looked wrong from the outside, while every hit quietly re-ran the whole report.
Truncated answers were being retried as-is. When the model spent its budget on thinking, the answer came back cut off, and nothing noticed, so the same work was simply run again. It gets recognised now.
A single mis-filled field no longer throws away the whole report. Roughly 13% of the time the model would put an emotion value into the attention field and the entire report would restart. Now only that one entry is dropped and the rest is kept.