A self-score from a sample overstated every line, one direction

The self-score, and what it missed

An editing pass scored its own sample work: the front matter, back matter, close, and one full part. It scored substance 9.0, actionability 9.5, enjoyment 6.5, and wrote PASSES.

Six blind readers then read the entire manuscript, roughly 80,266 words cover to close, each in their own fresh context with nothing but the manuscript and a brief. Same text, same book. Substance came back 8.5, actionability 7.5, enjoyment 4.0, and the book failed on enjoyment.

The gap ran one direction, every time

A full point overstated on substance. Two full points overstated on actionability. Two and a half points overstated on enjoyment, the line that decided the final verdict. The gap wasn't random noise running one way on one line and the other way on another; it ran the same direction on all three, every time, which is what a biased instrument looks like rather than an honest one that simply disagrees with itself occasionally.

Nothing about the original pass was careless. It read the parts it read closely and scored them honestly against what it found there. The failure sat one level up, in the decision to sample at all, and in letting the writer's own context anywhere near the scoring. Neither decision looks wrong from the inside. Both are wrong.

What done looks like

Nobody who drafted or edited a piece scores it. Every evaluation runs in a context with no memory of the drafting process, on the complete text, never a sample. A self-score, if one exists at all, is treated as unverified until an independent, blind, whole-text read confirms it.