humanize

Editing a good draft makes it worse. Rubric-guided model editing lifted below-median drafts on every measured dimension and pushed above-median drafts down: coherence −13.92, framing −12.06, both p<.001, across 45 abstract pairs and 869 keystroke-level edit logs. The first decision is never which tells to fix. It is whether this text should be rewritten at all.

Fix the writing, not the detection. Generated prose is worth fixing because it is vague, padded, and evasive. Detectors fire on post-training artifacts rather than on machine authorship, and they flag second-language and formulaic human writing at the same time. An edit whose purpose is a detector score is out of scope; every edit must improve the text for a reader.

Two failures carry equal weight. Under-editing leaves the tells and the text reads as generated. Over-editing flattens voice, deletes specifics, and produces a different slop — and it is the one the measurements above say is easy to cause.

Decide before editing

Read the whole text first. Nothing below is decidable from a search, because every one of these tells has a legitimate twin a few lines away.

Then judge the draft as a whole against three questions: does it hold one argument, do its paragraphs depend on each other, does it carry specifics that resist fabrication? Three clear yeses make a draft strong. A clear no to any one makes it weak. A question you cannot answer leaves it uncertain, and uncertain outranks weak: asking costs less than a rewrite that was never warranted.

The three questions are an inference, not a result. The measurement behind the gate split scientific abstracts by a rubric score, and nothing in it maps that split onto an article, a README, or a release note.

• Strong draft — review, do not rewrite. Mark the passages, name the mechanism behind each, propose the targeted edits, and leave everything unmarked untouched. Rewriting is the operation measured to cost coherence and framing here.
• Weak draft — rewrite. The same measurements show large gains on every dimension. Work the sweep below.
• Uncertain — mark, and ask. A draft you cannot place is a draft whose author knows something you do not.

Structural edits are what close the gap. In the same study, human editors changed almost nothing at any quality level, because they substituted and pruned locally while the global weakness survived. Reordering, merging, and cutting beat word swaps.