Ask the writer about negated setups, one at a time (1.1.0)

.claude-plugin/marketplace.json

@@ -9,7 +9,7 @@
     {
       "name": "quiron",
       "source": "./",
-      "version": "1.0.0",
+      "version": "1.1.0",
       "description": "AI humanizer skill: removes AI slop and the signs of AI writing from prose, then measures the rewrite against human baselines.",
       "author": {
         "name": "ilien",

CITATION.cff

@@ -9,7 +9,7 @@ type: software
 authors:
   - alias: ilien
     website: "https://github.com/ilien-dev"
-version: 1.0.0
+version: 1.1.0
 date-released: "2026-09-23"
 license: AGPL-3.0-only
 repository-code: "https://github.com/ilien-dev/quiron"

SKILL.md

@@ -160,7 +160,8 @@ Quirón progress:
 - [ ] 4. Meter: AI side fixed, then overshot
 - [ ] 5. Checklist: no FAIL, every READ and TELL ruled on
 - [ ] 6. check.sh converged (two clean runs in a row, with --source when rewriting)
-- [ ] 7. Final message follows the step 6 template (never "ready to publish")
+- [ ] 7. Review items asked one at a time, if the writer is in the conversation
+- [ ] 8. Final message follows the step 7 template (never "ready to publish")
 ```
 
 1. **Write or rewrite.** Apply the points above, the pattern index below, and the word
@@ -177,7 +178,8 @@ Quirón progress:
    have a median of 20 of 23 features in band. The goal is a clean checklist, not 23 of 23.
 5. **Converge** with `check.sh` (next section). When rewriting, pass the draft and any
    notes with `--source`; a number, date or link the sources do not have fails the run.
-6. **Report honestly**, with this template, in the user's language:
+6. **Review with the writer** (section below): `audit.py --review`, one question at a time.
+7. **Report honestly**, with this template, in the user's language:
 
    ```
    Result: <N> of 23 features in the human band; checklist <F> FAIL, <T> TELL; converged: yes/no.
@@ -250,6 +252,33 @@ any Title Case heading and any rate outside the band; it failed every one of 42
 human posts. The current rules pass 98% of them clean and still fail 86% of the Claude
 Sonnet and 93% of the GPT posts written on the same titles.
 
+## Review with the writer
+
+`scripts/audit.py --review [--json] FILE` lists sentences a reader may take for AI that
+the loop must not fix on its own. Today that is the negated setup: a short "X isn't Y."
+opening a paragraph or right before the point ("What surprised me most wasn't a
+number."). It is in 31% of assistant posts and 7% of human ones, but removing it inside
+the loop made rewrites read more AI to blind judges, because many of these sentences are
+ordinary ones (`references/numbers.md`). So the writer decides, sentence by sentence,
+and you leave these sentences exactly as they are while writing, rewriting and looping.
+
+Do this only when the writer is in the conversation, after `check.sh` converged; skip it
+in batch or headless runs. Ask about one item at a time, never as a list, and wait for
+the answer before the next. Each question carries:
+
+- the sentence, quoted, and which paragraph it is in;
+- why a reader may take it for AI: the reasons `--review` prints, in the writer's language;
+- your suggested change, written out: a rewrite that states the point, or deleting the
+  sentence when the paragraph works without it;
+- four choices: **Keep**, **Use the suggestion**, **Delete**, **Write my own**.
+
+Use the host's question tool when it has one (in Claude Code, AskUserQuestion: the first
+three as options with the suggestion in the preview, and its free-text answer as "Write
+my own", used word for word); otherwise ask in plain text with the same four choices.
+Apply each answer before asking the next. A suggestion adds no fact and is not another
+negated setup. After the last answer, if the text changed, run `check.sh` once more and
+report the result; never undo the writer's choice to clear a TELL.
+
 ## Pattern index
 
 What `audit.py` checks. The state is what it reports. Full entries, with measured rates

references/numbers.md

@@ -10,6 +10,7 @@ The bands, the AI directions, what the whole apparatus does on held-out texts, h
 - Four counter-intuitive results
 - What the apparatus does on held-out texts
 - What the loop does to a text
+- Checks measured and not shipped
 - Calibrating for another register
 - Part D: what this does not do
 
@@ -155,6 +156,38 @@ python3 scripts/evaluate.py --human CORPUS --ai-train eval/ai/blog/train \
     opus=eval/ai/blog/test-opus sonnet=eval/ai/blog/test-sonnet gpt=eval/ai/blog/test-gpt
 ```
 
+### Checks measured and not shipped
+
+Each of these was measured with the same held-out split and left out, so it is not
+proposed again without new evidence. Human corpus: 176 dev.to posts (44 held out), rebuilt
+in September 2026.
+
+- **Stock phrases** (2 to 4 words, chosen on the training split). Phrases in at most 4% of
+  human and at least 10% of assistant training posts: 41% of held-out human posts carry
+  three or more, so the list flags people. Phrases in no human training post: AUC 0.53 to
+  0.57 against held-out assistant posts, close to chance. Rewrites by the skill already sit
+  at the human level on both lists (AUC 0.45 to 0.50).
+- **Length as verbosity.** Opus 5.5 and Sonnet 5 asked for a dev.to post with no length
+  wrote 1.41 and 1.27 times the words of the human post on the same title (AUC 0.67 and
+  0.59, 25 posts). The extra is code and headings: prose alone was shorter than the
+  human's (0.70 times), with shorter sentences. Headings are already a band; a length rule
+  has no human text to compare a single post against.
+- **Negated setup (A1c)**: a sentence of 12 words or fewer that only denies something,
+  opening a paragraph or followed by the reframe ("What surprised me most wasn't a
+  number."). As a checklist item it looked right: 7% of held-out human posts, 31% of
+  assistant posts, held-out humans still 0% T1 FAIL, T1 FAIL on GPT posts from 7% to 36%.
+  End to end it made rewrites worse. 20 titles (8 fresh), 52 rewrites per version, two
+  fresh Opus judges each: mean P(AI) 43.6 without the check and 46.4 with it, higher on 13
+  of 20 titles (Wilcoxon p = 0.036); on the fresh titles 34.6 against 40.5. The countable
+  rates did not move. Many of the sentences it removed were ordinary ones from the
+  author's notes ("The page isn't finished yet."), and the judges' reasons for the
+  rewrites with the check were "too even" and "uniformly polished". It is kept as a review item
+  the writer rules on (`audit.py --review`), outside the loop, T1 and `check.sh`. With
+  that version of the skill, 40 more rewrites of the same 20 titles (two per title,
+  headless, so no review questions) gave mean P(AI) 43.7 against 43.6 without it (+1.3
+  per title, 95% CI -2.0 to +4.7, Wilcoxon p = 0.42), and the countable rates did not
+  move. Negated setups were left in 5 of 40 rewrites, against 11 of 52 before.
+
 ### Calibrating for another register
 
 **Calibrating for another register.** Academic prose, news, email and other languages

references/patterns.md

@@ -47,12 +47,17 @@ State in brackets is what `audit.py` reports.
 
 **Watch for:** it's not X, it's Y; not just / not only / not merely X, but Y; X rather
 than Y; the contrast split across two sentences ("This does not mean X. It means Y.");
-the clipped negative tail ("…, no guessing", "That's a favor, not a liability.").
+the clipped negative tail ("…, no guessing", "That's a favor, not a liability."); the
+negated setup, a short sentence that only denies something, opening a paragraph or right
+before the point ("What surprised me most wasn't a number.").
 **Measured:** the family is in 38% of assistant training posts and 18% of human ones; "rather
 than" alone was in 35% and 8% on the training split. "It's not X, it's Y" is in 5% of
 human posts and 9% of assistant ones, so it cannot fail on sight. Across 22 models the
 slop-score benchmark (Paech 2025) finds the construction at 2 to 13 times the human rate;
-the rate, not a single use, is the tell.
+the rate, not a single use, is the tell. The negated setup is in 7% of held-out human posts
+and 31% of assistant ones, but fixing it inside the loop made rewrites read more AI to
+blind judges, so it is a review item only: `audit.py --review` lists it and the writer
+decides (SKILL.md, "Review with the writer"). Blog bands only.
 **Why it fails:** the negative half names something nobody claimed, so the positive half
 sounds larger. It adds weight without adding a claim.
 **Instead:** state the point. Keep the contrast only when the negative half corrects a

scripts/audit.py

@@ -27,6 +27,9 @@
   audit.py --json FILE
   audit.py --sample PATH FILE  a file or folder of the writer's own texts: a TELL that
                                fires there too is their habit and is not counted
+  audit.py --review [--json] FILE
+                           sentences for the writer to decide on, one at a time; never
+                           part of the checklist, T1 or check.sh, and the exit code is 0
 """
 import json, os, re, subprocess, sys
 
@@ -196,6 +199,41 @@ def _thresholds():
         c for c in CHECKS if c[0] in ("C1", "C2", "T1", "V1", "V2")]
 
 SENT = re.compile(r"(?<=[.!?])\s+")
+
+# Review items: patterns a reader notices that the rewrite loop must not act on alone.
+# A1c (negated setup) is in 7% of held-out human posts and 31% of assistant posts, but
+# fixing it inside the loop made rewrites read more AI to blind judges (numbers.md,
+# "Checks measured and not shipped"), so only the writer decides, sentence by sentence.
+NEGSET = re.compile(r"\b(?:isn't|wasn't|aren't|weren't|is not|was not|are not|were not)\b", re.I)
+NEGSET_SKIP = re.compile(r"\b(?:but|because|if|when|unless|so)\b", re.I)
+REFRAME = re.compile(r"(?:it's|it is|it was|that's|that is|this is|what |the (?:real|actual|point|answer)|"
+                     r"instead|rather)", re.I)
+NEGSET_WHY = [
+    "It denies something nobody said, so the sentence after it sounds bigger than it is.",
+    "It sits where the point should be: {where}.",
+    "A short 'X isn't Y.' sentence like this is in 31% of assistant blog posts and 7% of "
+    "human ones (held-out dev.to posts).",
+]
+
+
+def review(path):
+    """Negated setups, with the reasons a reader may take them for AI. Blog bands only."""
+    if lexicon_name() != "ai-lean.txt":
+        return []
+    items = []
+    for n, para in enumerate(re.split(r"\n\s*\n", prose(open(path, encoding="utf-8").read()))):
+        ss = [s_.strip() for s_ in SENT.split(para.replace("\n", " ")) if s_.strip()]
+        for i, s_ in enumerate(ss):
+            nxt = ss[i + 1] if i + 1 < len(ss) else ""
+            opens, reframed = i == 0, bool(REFRAME.match(nxt))
+            if (len(s_.split()) <= 12 and s_.endswith(".") and NEGSET.search(s_)
+                    and not NEGSET_SKIP.search(s_) and (opens or reframed)):
+                where = ("it opens the paragraph, before the real point" if opens
+                         else "right before the sentence that makes the real point")
+                items.append({"id": "A1c", "name": "negated setup", "paragraph": n + 1,
+                              "sentence": s_, "next": nxt,
+                              "why": [w.format(where=where) for w in NEGSET_WHY]})
+    return items
 MINOR = set("a an the and or but of to in on for with vs at by from is as into via".split())
 
 
@@ -348,6 +386,19 @@ def main():
     missing = [p for p in paths if not os.path.isfile(p)]
     if missing:
         sys.exit(f"no such file: {', '.join(missing)}")
+    if "--review" in sys.argv:
+        for path in paths:
+            items = review(path)
+            if as_json:
+                print(json.dumps({"file": path, "items": items}, indent=1))
+                continue
+            print(f"\n{os.path.basename(path)}: {len(items)} to review")
+            for k, it in enumerate(items, 1):
+                print(f"\n{k}. {it['id']} {it['name']}, paragraph {it['paragraph']}")
+                print(f"   \"{it['sentence']}\"")
+                for w in it["why"]:
+                    print(f"   - {w}")
+        return 0
     fails = 0
     for path in paths:
         res = run(path, samples)