anti-sloplisted

# anti-slop Slop is not a vocabulary problem. It is manufactured upstream: alignment training collapses output toward the statistical mean (mode collapse, driven by typicality bias in preference data — arXiv 2510.01171), single-sample generation returns the most typical completion, and a context starved of specifics can only produce the average of everything. Banned-word lists treat the symptom. At adoption scale they mint a new house style — the zero-em-dash, no-adverbs voice is already its own tell, and Wikipedia's own tells-catalog maintainers warn that scrubbing signs "could just make detection harder." **The job: make the output sound like a specific person with something specific to say.** Both halves are mandatory. This skill intervenes at four layers and measures whether it worked. It is a writing-quality protocol, not a detector-evasion tool. --- ## Modes | Mode | When | Run | |---|---|---| | **DRAFT** | writing new text | Layers 1→2→3→4, ship with score | | **AUDIT** | handed existing text | Layers 3→4, report score before/after + fix list | --- ## The contract — verify before generating anything Two inputs separate writing from slop. Check both first: **1. A bound voice.** In priority order: - an explicit voice spec (voice file, brand guide), or - 5+ writing samples from the author → build a voice card on the fly (`references/voice-binding.md`), or - a named register the user states ("tired senior engineer in Slack", "group-chat shitposter who knows financ