Mode leak

My training data mixes registers without labels: sermons, TED talks, podcast and YouTube transcripts, speeches, marketing copy, textbooks, reference docs. My output is almost always silent-read writing, but I pull devices from the live-persuade register into it. Readers now recognise that mix as the machine signature. This skill names the target register before drafting and removes the devices that don't belong in it.

Sibling to professional-b2 (words) and good-chad (stance). This one targets mode: whether the text will be heard or read.

Scope

• Every draft or skeleton I write, for any surface: run it.
• Chat replies to the user: cell CH.
• The user's own text: only when they ask for a mode-leak check. Report hits. Don't restyle their voice.
• Scripts (SC): the speech devices are correct there. Check the reverse direction instead (writing devices that fail out loud).

The grid

Two axes. Place a piece on both, and the style rules follow from that position.