Almost none of it was facts

We assumed the edits would be corrections. Grounding the model properly had been most of our work, so we expected the remaining errors to be factual. They were not. Factual edits were a small minority, and most of those were on pieces where the source material was genuinely ambiguous.

The overwhelming majority of edits were tonal, and they clustered tightly around a handful of recognisable habits.

The four habits

Sorted by how often a human deleted them.

• Throat-clearing openers — a sentence that announces the topic before saying anything about it
• Hedging that the source did not contain, added because the model was uncertain and the prose was not allowed to be
• False enthusiasm, which shows up reliably when a model has nothing specific to say
• Summary paragraphs that restate the piece instead of ending it

Why models reach for them

All four are safe. They are what you produce when you are optimising to not be wrong rather than to be useful. A hedge is never incorrect. An opener that restates the title is never off topic.

That framing helped more than any individual fix. We stopped treating these as style bugs and started treating them as a symptom of the model having insufficient specific material to work with.

What we changed

Two things. The prompt layer now refuses to generate a piece when the grounding is too thin, rather than padding. And the voice spec gained explicit rules against all four, with examples.

Throat-clearing openers dropped the most. False enthusiasm is the stubborn one, because it correlates with genuinely thin source material, and the real fix there is better inputs.