An agent writing in your voice does not only sample the median of the internet. It\nsamples the median of you, from the pieces and rules and memory files in its context,\nand it regresses toward your most frequent choices because those are the highest\nprobability tokens in your corpus. A word that was a choice at one use comes back at\nfive. Then the fifth use teaches the next draft, and the tic becomes the style.
\nGeneric ban lists cannot catch this, because the words are yours and they are not on\nanyone's list. Measure instead.
\npython3 scripts/census.py PATH/TO/YOUR/WRITING --out voice-baseline.md\npython3 scripts/census.py PATH/TO/YOUR/WRITING --check DRAFT.md\nThe census walks a directory of your own finished pieces and reports word counts with\na per-file rate, sentence-shape counts, sentence openers, and heading first words.\n--check compares one draft against those rates and prints only what sits above them.\nExit code 1 when something is above baseline, so it fits a pre-publish script.
Keep the output file next to your rules so the agent reads it before it writes.
\nWords. A watchlist of praise and emphasis words, counted with word boundaries and\nreported per file. Read the per-file rate: a word used twice per file is a habit, and\nthe same word at eight in one draft is the draft's problem. Edit WATCH in the\nscript to add your own suspects; the point is that the list ends up specific to you.
Paragraph-ending verdict sentences. \"The fix was one line.\" A concrete noun, a\npast-tense verb, a short predicate, parked at the end of a paragraph so it lands like\na ruling. One per piece at most, and only when it carries a fact the paragraph has not\nalready delivered.
\nKicker closers. \"That is the point.\" \"That is all a title is for.\" A closing line\nthat restates the piece as an aphorism. The cap is zero.
\nEndings. How many pieces close on a question to the reader. Most of them means the\nquestion has become a formula.
\nSentence openers and heading first words. These are where a personal tic hides\nbest, because each instance reads fine. If a third of your headings begin with the\nsame word, you have a template. In one real 27-file archive the census found 53\nheadings starting with \"The\", the exact shape people name as an AI tell, learned from\nthe author's own archive.
\nThe census counts; it does not judge. For each flagged word or shape, decide whether\nit is your voice or your mode:
\nFix the shapes by cap rather than by ban: one verdict sentence, zero kickers, no\ndefault question ending, one narrative heading pattern per piece.
\nRun the census again after roughly ten new pieces. A tic you fixed should fall. A new\none will have appeared, usually the word you reached for while avoiding the last one,\nwhich is the same substitution the model makes when a single word is banned.
\nPair with slop-gate for the universal tells and no-ai-slop for the structural ones\na regular expression cannot see.