GlyphGuard

GlyphGuard provides an interface for inspecting documents, pasting text, inspecting suspicious Unicode and AI artefacts in real time, tuning what gets flagged, and exporting the full report when needed. It works locally, so you can review what was found and decide what matters without sending your text anywhere.

What it does

GlyphGuard is a small local inspection tool for text and documents. It shows suspicious Unicode, hidden characters, unusual punctuation, private-use markers, leaked citation plumbing, raw tool metadata and repeated generated-writing patterns. It does not prove authorship and it does not send the text to an external detector. It gives you a structured set of things to check.

Why use it

Some problems are easy to see, such as a paragraph that is full of generic generated prose or a copied tool marker left in the final text. Other problems are much easier to miss: zero-width characters inside words, non-breaking spaces that break search, direction controls that change display order, private-use citation markers, malformed metadata, or odd Unicode that causes problems in diffs, moderation tools, filenames and parsers.

GlyphGuard is useful for:

• checking pasted text before publishing it;
• reviewing documents, wiki edits, Markdown, logs or support text without using an online detector;
• finding hidden or unusual Unicode that is not obvious in normal editors;
• spotting leaked AI/tool artefacts, citation placeholders and raw metadata;
• comparing weak writing-style hints with stronger evidence such as copied tool output;
• exporting a simple report when you need to show what was found and where.

It is a small tool I made for practical review work, not a polished detection service. The point is to make the suspicious bits visible so you can decide what matters.

GlyphGuard does not prove that a passage was written by AI. Its heuristic rules highlight review signals: unusual glyphs, leaked tool markers, copied citation plumbing, generic generated-prose patterns and other artefacts worth checking.

Features

This is the high-level overview. Later sections explain how to interpret the results, run the scorer, add samples and update rules.

GlyphGuard is built with Tkinter and standard Python libraries. Report generation uses matplotlib. It runs locally on Windows, macOS and Linux.

Unicode and glyph inspection

Paste text or load a supported file and GlyphGuard lists non-ASCII characters with their code point, Unicode name, category, concern level, count and positions. It covers invisible controls, non-breaking spaces, bracket alternatives, dash and quote variants, fullwidth forms, private-use markers and other derived Unicode oddities.

This is useful for AI review because some AI tools leak private-use markers or odd citation characters. It is also useful outside AI detection because hidden characters can be used for prompt injection, confusing diffs, broken search, bad copy/paste, parser failures, misleading identifiers, or text that displays differently from how it is stored.

Useful checks include:

• zero-width spaces, joiners and non-joiners hidden inside ordinary words;
• direction controls and invisible formatting characters;
• curly quotes, long dashes, fullwidth letters or strange brackets before pasting into strict systems;
• private-use characters left behind by citation or rendering tools;
• replacement suggestions for making a clean copy of the text.

The UI also uses semantic badges. Character rows show category badges, and holistic rules are tagged by evidence type such as LLM Style, Structure, Evidence Gap, Citation Leak, Metadata Leak, API Leak, Punctuation Pattern and vendor/model tags where the rule is specific enough. Dense badge rows collapse behind a +N expander when needed.

AI artefact detection

GlyphGuard uses a curated set of regular expressions to catch vocabulary, phrasing, structure, citation leaks, tool artefacts and API metadata that are often worth reviewing in generated or copy-pasted AI output.

A compact example of the sort of thing it is meant to catch:

The source says this is supported by [12†L44-L51] and turn0file3.

The exact rendered marker changes by tool, but this is the shape of the problem: citation plumbing that should have been converted into normal prose, a normal footnote, or a normal link before the text was published.

Rule families cover OpenAI browser and Deep Research citations, private-use citation markers, source IDs, Harmony tokens, File Search metadata, Claude citation metadata, Gemini grounding metadata and Perplexity/Sonar search metadata.

Holistic writing-pattern rules

Holistic rules are broader writing-pattern checks. They look for repeated generated-prose habits, formulaic structures, vague attribution, suspiciously generic framing, evidence gaps and punctuation patterns.

These rules are scored because a single weak phrase is not very meaningful. Repeated matches across a short text are more interesting than one isolated match in a long document. The score is a review priority, not a probability and not an authorship verdict.

Each holistic rule can include static examples, live matches from the scanned text, source references, semantic tags, positions, severity and a suggested review action. Examples explain the rule. Matches are the actual snippets found in your text.

Document and pasted-text workflow

You can paste text directly into the input pane or import a file. Supported input covers plain text, Markdown, logs, CSV/TSV, JSON, YAML, TOML, INI, source-code files, HTML/XML/SVG, DOCX, ODT, RTF and text-based PDFs.

The application scans pasted text and loaded document content as you work, updating counts and flags immediately without server calls. You can collapse noisy panels, click position pills to jump back to the source text, hover source badges for provenance, and adjust character detection and replacement behaviour in preferences.

For plain text sources, GlyphGuard can save cleaned text back to the original file or save a clean copy. For document formats such as DOCX, ODT, RTF and PDF, extraction is read-only; use Save as or Save clean copy.

Reports and exports

GlyphGuard can export detailed PDF and plain-text reports. Reports include summary scores, glyph detections, holistic matches, examples, sources, positions and charts.

The headless scorer can also scan files and folders without opening the Tkinter UI.

Sample corpus and scoring

The samples/ folder currently contains samples/wikipedia_signs_ai_writing/, a bundled set of 198 text samples extracted from Wikipedia's "Signs of AI writing" page. The set contains 175 quote examples and 23 preformatted examples.

The corpus is for detector tuning, regression checks and quick scoring experiments. It is not a claim that every sample is AI-generated.

Quick start

Requirements

• Python 3.8 or newer.
• matplotlib for PDF report generation.
• Tkinter for the desktop UI. It is bundled with many Python installs. On some Linux distributions you may need to install it separately, for example with python3-tk.
• pytest if you want to run tests.

Install

From a fresh clone:

git clone <repo-url>
cd glyphguard
python -m pip install matplotlib

For development:

python -m pip install matplotlib pytest

There is no packaging wrapper at the moment. Run the files from the repository root.

Run the app

python glyphguard.py

The app opens a resizable desktop window. Paste text into the input pane or use Import to load a document. If the window does not open on Linux, check that Tkinter is installed.

Run the sample scorer

Default run, using the bundled sample corpus when present:

python score_samples.py

Score the bundled corpus explicitly:

python score_samples.py samples/wikipedia_signs_ai_writing --out results

Score a specific file or folder:

python score_samples.py README.md docs --out results

Choose extensions when scoring a mixed folder:

python score_samples.py docs --extensions .txt,.md,.html,.docx,.pdf --out results

Run a quick smoke test:

python score_samples.py samples/wikipedia_signs_ai_writing --limit 10 --out /tmp/glyphguard_score_smoke

The older --samples option is still accepted as an alias for one input file or folder, but positional paths are the normal form now.

Typical workflow

1. Launch GlyphGuard.
2. Paste text, or click Import and choose a file.
3. Review the score cards and detection lists.
4. Expand rows to see positions, examples, tags and sources.
5. Use the clean-text preview to check replacements.
6. Export a PDF or text report if you need to share the findings.

Understanding results

GlyphGuard results should be read as review prompts. Strong artefacts such as leaked private-use markers or raw citation metadata are different from weak style hints such as a generic transition. The useful question is not "did the tool accuse this text?" but "what exactly did it find, how strong is that signal, and does the surrounding text support that concern?"

A good review usually works in this order:

1. Check very-high and high severity artefacts first.
2. Check whether the match is a hard leak, a hidden character, or only a writing-style hint.
3. Use positions and snippets to inspect the original context.
4. Treat repeated weak signals as more interesting than isolated weak signals.
5. Keep false positives and false negatives in mind.

Severity levels

GlyphGuard uses practical concern levels:

• None - visible but not normally a problem, or kept as a no-concern review signal.
• Low - weak hint or low-risk formatting oddity.
• Medium - worth checking, especially if repeated.
• High - likely to matter, such as malformed markup, suspicious metadata or stronger style patterns.
• Very high - serious glyph concern or mapped critical artefact.
• Critical - internal holistic-rule severity for hard tool leaks, prompt leaks, raw metadata or placeholder artefacts. In the UI this maps into the very-high concern bucket.

Rule matches

Rules have static examples in the code and reports. Those examples are not evidence from the current scan; they are there to show what the rule is looking for.

Matches are different. A match is text found in the document you scanned, with positions where possible. When reviewing a rule row, use examples to understand the rule and matches to inspect the actual evidence.

Glyph matches

Glyph rows show the character code point, Unicode name, category, concern level, count, positions and possible replacement. Unknown non-ASCII characters can still be derived from Python's Unicode metadata and shown for review even when they are not in the curated list.

Replacement suggestions are deliberately conservative. Some characters should be removed, some should be replaced with plain ASCII, and some should be kept because they are normal in the right context. The clean-text preview is there so you can check the result before saving anything.

Tags, categories and sources

Holistic rules use semantic tags such as LLM Style, Structure, Evidence Gap, Citation Leak, Metadata Leak, API Leak, Punctuation Pattern, OpenAI, Claude, Gemini and Perplexity. Character rows use category badges such as space, bracket, punctuation, invisible format and safe symbol.

Sources are provenance, not proof. A source badge explains where the idea for a rule came from or which public documentation describes that artefact shape. Hover text in the UI and the exported reports include the source name and description so the rule is easier to audit later.

What GlyphGuard does not prove

GlyphGuard is not an authorship detector. It will not prove a person used AI, and it will not prove a person did not use AI.

False positives are expected. Humans write generic transitions. Humans use dashes. Humans paste weird Unicode by accident. False negatives are also expected. AI output can be edited, paraphrased or generated without obvious tool leaks.

Treat hard artefacts differently from weak style hints. A leaked private-use citation marker is strong copy/paste evidence. A generic transition is only a prompt to look closer.

Scores

Scores are for prioritising review, not judging authorship.

• Raw score adds up weighted matches.
• Contextual score adjusts the raw evidence using density and spread across the text.
• Character score covers glyph-level findings.
• Holistic score covers writing-pattern and tool-artefact rules.
• Total score combines both.

A short text with one hard citation leak can still score high. A long document with a few mild signals should not be treated the same way. That is why the contextual score exists.

Score bands

Family	Clean	Trace	Low	Medium	High	Very high
Glyph	0	>0 and < 2	>= 2 and < 6	>= 6 and < 20	>= 20 and < 50	>= 50
Holistic	0	>0 and < 1.5	>= 1.5 and < 4	>= 4 and < 10	>= 10 and < 20	>= 20
Total	0	>0 and < 3	>= 3 and < 10	>= 10 and < 30	>= 30 and < 70	>= 70

Raw vs contextual scores

Raw score is the direct evidence total. For glyphs, each detected character occurrence contributes a concern weight. For holistic rules, each matched pattern contributes the pattern weight from HOLISTIC_AI_PATTERN_SPECS.

Contextual score keeps that raw evidence as the anchor, then adjusts it using density and spread. Dense evidence in a short text is more concerning than the same raw score scattered through a long text. Evidence spread across many sentences or paragraphs is also more concerning than one localised match.

This is still a heuristic. It is meant to sort review priority, not produce a probability.

Contextual score maths

The maths uses smooth saturation rather than hard cut-offs. The main idea is that ratios such as "matched words divided by total words" should increase concern quickly at first, then flatten out instead of growing forever. The function used for that is based on exponential decay: Exponential decay. The weighted combination is just a weighted average: Weighted arithmetic mean.

Smooth saturation

For a ratio x and a scale value s, GlyphGuard uses:

$$ S(x, s) = 1 - e^{-(x / s)} $$

Where:

• x is the observed ratio, such as matched words divided by all words.
• s is the scale that controls how quickly the curve approaches 1.
• S(x, s) is the bounded factor used by the contextual multiplier.

Small values of x still count. Very large values do not explode the score.

Glyph raw score

For glyphs, each concern level has a weight:

Concern	Weight
None	0
Low	0.5
Medium	1.0
High	2.0
Very high	3.0

The raw glyph score is:

$$ R_g = \sum_i count_i \times weight_i $$

Glyph contextual score

Glyph context uses character density and line spread:

$$ d_g = \frac{flagged\ occurrences}{total\ characters} $$

$$ l_g = \frac{matched\ lines}{total\ lines} $$

Those ratios are converted into saturation factors:

$$ D_g = S(d_g, 0.01) $$

$$ L_g = S(l_g, 0.20) $$

The glyph context factor is:

$$ C_g = 0.80D_g + 0.20L_g $$

The final glyph multiplier and contextual glyph score are:

$$ M_g = 0.70 + 1.30C_g $$

$$ Score_g = R_g \times M_g $$

Holistic raw score

For holistic matches, raw score is the sum of matched pattern weights:

$$ R_h = \sum_j weight_j $$

The scorer only counts matches that overlap prose blocks or blocks with matches. This avoids treating every bit of boilerplate or non-prose file content as equally meaningful.

Holistic contextual score

Holistic context uses word density, sentence spread and paragraph spread:

$$ d_w = \frac{matched\ words}{total\ prose\ words} $$

$$ s_h = \frac{matched\ sentences}{total\ sentences} $$

$$ p_h = \frac{matched\ paragraphs}{total\ paragraphs} $$

Word density uses saturation directly:

$$ W_h = S(d_w, 0.05) $$

Sentence and paragraph spread also get a small-sample confidence factor so one sentence or one paragraph cannot look stronger than it is:

$$ Conf(n, k) = \frac{n}{n + k} $$

$$ S_h = S(s_h, 0.30) \times Conf(total\ sentences, 8.0) $$

$$ P_h = S(p_h, 0.60) \times Conf(total\ paragraphs, 5.0) $$

The holistic context factor is weighted towards word density because paragraph counts can be volatile in short text:

$$ C_h = 0.55W_h + 0.30S_h + 0.15P_h $$

The final multiplier and contextual holistic score are:

$$ M_h = 0.50 + 1.50C_h $$

$$ Score_h = R_h \times M_h $$

Total score

The total score just adds the two families:

$$ R_t = R_g + R_h $$

$$ Score_t = Score_g + Score_h $$

The UI shows contextual score first and raw score in brackets, because contextual score is usually the better review priority.

Detection rule families

The detection rules are defined in glyphguard.py so the app remains self-contained, might expand it in future or allow loading from somewhere in future maybe.

Unicode and hidden-character rules

Glyph-level detections live in TRACKED_CHAR_SPECS until I think of a better name. The app also derives fallback metadata from Python's Unicode database for non-ASCII characters that are not in the curated list. Rules cover invisible controls, private-use markers, bracket variants, whitespace variants, dash and quote variants, fullwidth characters, replacement suggestions and derived unknown Unicode.

The table lists code points and escaped examples rather than inserting every raw character directly into the README. That keeps invisible/private-use characters from hiding inside the documentation itself.

Tracked character table (127 entries)

Code	Visibility	Concern	Categories	Description	Kind	Example	Replacement	Keyboard	HTML/raw
U+0009	Visible	Low concern	Tracked, Space, Control	Tab	Tracked character	word ? word	? 4 spaces	Tab	-
U+000A	Visible	No concern	Tracked, Space, Control	Line Feed (LF)	Tracked character	line1\\nline2	keep in double	Enter	-
U+000C	Visible	High concern	Tracked, Control	Form Feed	Tracked character	-	remove	-	-
U+000D	Visible	No concern	Tracked, Space, Control	Carriage Return (CR)	Tracked character	-	keep in double	Enter	-
U+001C	Invisible	Very high concern	Tracked, Invisible format, Control	File Separator	Tracked character	-	remove	-	-
U+0020	Visible	No concern	Tracked, Keyboard, Space	Regular Space	Tracked character	word word	keep	Spacebar	-
U+00A0	Visible	Medium concern	Tracked, Space	No-Break Space	Tracked character	word word	? space	-	-
U+00A2	Visible	No concern	Tracked, Safe symbol, Symbol	Cent Sign	Common safe symbol	50\\xa2	keep	-	\\xa2
U+00A3	Visible	No concern	Tracked, Keyboard, Safe symbol, Symbol, GB language char	Pound Sign	Common safe symbol	\\xa310	keep	\\xa3	\\xa3
U+00A4	Visible	No concern	Tracked, Safe symbol, Symbol	Currency Sign	Common safe symbol	\\xa4	keep	-	\\xa4
U+00A5	Visible	No concern	Tracked, Safe symbol, Symbol	Yen Sign	Common safe symbol	\\xa51000	keep	-	\\xa5
U+00A7	Visible	No concern	Tracked, Safe symbol, Punctuation	Section Sign	Common safe symbol	\\xa7 1	keep	-	\\xa7
U+00A9	Visible	No concern	Tracked, Safe symbol, Symbol	Copyright Sign	Common safe symbol	\\xa9 2026	keep	-	\\xa9
U+00AD	Invisible	Very high concern	Tracked, Invisible format, Control	Soft Hyphen	Tracked character	-	remove	-	\\xad
U+00AE	Visible	No concern	Tracked, Safe symbol, Symbol	Registered Sign	Common safe symbol	Name\\xae	keep	-	\\xae
U+00B0	Visible	No concern	Tracked, Safe symbol, Symbol	Degree Sign	Common safe symbol	21\\xb0C	keep	-	\\xb0
U+00B6	Visible	No concern	Tracked, Safe symbol, Punctuation	Pilcrow Sign	Common safe symbol	\\xb6	keep	-	\\xb6
U+0192	Visible	No concern	Tracked, Safe symbol, Letter, International char	Latin Small Letter F With Hook / Florin	Common safe symbol	\\u019210	keep	-	\\u0192
U+034F	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Grapheme Joiner	Tracked character	-	remove	-	\\u034f
U+115F	Visible	Low concern	Tracked, Letter, International char	Hangul Choseong Filler	Tracked character	-	remove	-	\\u115f
U+1160	Visible	Low concern	Tracked, Letter, International char	Hangul Jungseong Filler	Tracked character	-	remove	-	\\u1160
U+1680	Visible	Medium concern	Tracked, Space	Ogham Space Mark	Tracked character	-	? space	-	-
U+17B4	Visible	Low concern	Tracked, Mark, International char	Khmer Vowel Inherent AQ	Tracked character	-	remove	-	\\u17b4
U+17B5	Visible	Low concern	Tracked, Mark, International char	Khmer Vowel Inherent AA	Tracked character	-	remove	-	\\u17b5
U+180B	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Mongolian VS-1	Tracked character	-	remove	-	\\u180b
U+180C	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Mongolian VS-2	Tracked character	-	remove	-	\\u180c
U+180D	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Mongolian VS-3	Tracked character	-	remove	-	\\u180d
U+180E	Visible	Medium concern	Tracked, Space, Control	Mongolian Vowel Sep.	Tracked character	-	? space	-	\\u180e
U+2000	Visible	Medium concern	Tracked, Space	En Quad	Tracked character	-	? space	-	-
U+2001	Visible	Medium concern	Tracked, Space	Em Quad	Tracked character	-	? space	-	-
U+2002	Visible	Medium concern	Tracked, Space	En Space	Tracked character	-	? space	-	-
U+2003	Visible	Medium concern	Tracked, Space	Em Space	Tracked character	-	? space	-	-
U+2004	Visible	Medium concern	Tracked, Space	Three-Per-Em Space	Tracked character	-	? space	-	-
U+2005	Visible	Medium concern	Tracked, Space	Four-Per-Em Space	Tracked character	-	? space	-	-
U+2006	Visible	Medium concern	Tracked, Space	Six-Per-Em Space	Tracked character	-	? space	-	-
U+2007	Visible	Medium concern	Tracked, Space	Figure Space	Tracked character	-	? space	-	-
U+2008	Visible	Medium concern	Tracked, Space	Punctuation Space	Tracked character	-	? space	-	-
U+2009	Visible	Medium concern	Tracked, Space	Thin Space	Tracked character	-	? space	-	-
U+200A	Visible	Medium concern	Tracked, Space	Hair Space	Tracked character	-	? space	-	-
U+200B	Invisible	Very high concern	Tracked, Invisible format, Control	Zero-Width Space	Tracked character	-	? space	-	-
U+200C	Invisible	Very high concern	Tracked, Invisible format, Control	Zero-Width Non-Joiner	Tracked character	-	remove	-	\\u200c
U+200D	Invisible	Very high concern	Tracked, Invisible format, Control	Zero-Width Joiner	Tracked character	-	remove	-	\\u200d
U+200E	Invisible	Very high concern	Tracked, Invisible format, Control	Left-To-Right Mark	Tracked character	-	remove	-	\\u200e
U+200F	Invisible	Very high concern	Tracked, Invisible format, Control	Right-To-Left Mark	Tracked character	-	remove	-	\\u200f
U+2013	Visible	Medium concern	Tracked, Punctuation	En Dash	Tracked character	pages 1\\u201310	-	-	\\u2013
U+2014	Visible	Medium concern	Tracked, Punctuation	Em Dash	Tracked character	word\\u2014word	-- or -	-	\\u2014
U+2018	Visible	Medium concern	Tracked, Punctuation	Left Single Quote	Tracked character	'hello	'	-	\\u2018
U+2019	Visible	Medium concern	Tracked, Punctuation	Right Single Quote	Tracked character	don't	'	-	\\u2019
U+201C	Visible	Medium concern	Tracked, Punctuation	Left Double Quote	Tracked character	"hello	"	-	\\u201c
U+201D	Visible	Medium concern	Tracked, Punctuation	Right Double Quote	Tracked character	hello"	"	-	\\u201d
U+2026	Visible	Medium concern	Tracked, Punctuation	Horizontal Ellipsis	Tracked character	wait...	...	-	\\u2026
U+2028	Invisible	Very high concern	Tracked, Invisible format	Line Separator	Tracked character	-	? \n	-	-
U+2029	Invisible	Very high concern	Tracked, Invisible format	Paragraph Separator	Tracked character	-	? \n\n	-	-
U+202A	Invisible	Very high concern	Tracked, Invisible format, Control	LTR Embedding	Tracked character	-	remove	-	\\u202a
U+202B	Invisible	Very high concern	Tracked, Invisible format, Control	RTL Embedding	Tracked character	-	remove	-	\\u202b
U+202C	Invisible	Very high concern	Tracked, Invisible format, Control	Pop Directional Fmt	Tracked character	-	remove	-	\\u202c
U+202D	Invisible	Very high concern	Tracked, Invisible format, Control	LTR Override	Tracked character	-	remove	-	\\u202d
U+202E	Invisible	Very high concern	Tracked, Invisible format, Control	RTL Override	Tracked character	-	remove	-	\\u202e
U+202F	Visible	Medium concern	Tracked, Space	Narrow NBSP	Tracked character	-	? space	-	-
U+2045	Visible	Medium concern	Tracked, Bracket, Punctuation	Left Square Bracket With Quill	Bracket alternative	\\u2045text	[	[	\\u2045
U+2046	Visible	Medium concern	Tracked, Bracket, Punctuation	Right Square Bracket With Quill	Bracket alternative	text\\u2046	]	]	\\u2046
U+205F	Visible	Medium concern	Tracked, Space	Math Space	Tracked character	-	? space	-	-
U+2060	Invisible	Very high concern	Tracked, Invisible format, Control	Word Joiner	Tracked character	-	remove	-	\\u2060
U+2061	Invisible	Very high concern	Tracked, Invisible format, Control	Function Application	Tracked character	-	remove	-	\\u2061
U+2062	Invisible	Very high concern	Tracked, Invisible format, Control	Invisible Times	Tracked character	-	? "x"	-	\\u2062
U+2063	Invisible	Very high concern	Tracked, Invisible format, Control	Invisible Separator	Tracked character	-	? ","	-	\\u2063
U+2064	Invisible	Very high concern	Tracked, Invisible format, Control	Invisible Plus	Tracked character	-	? "+"	-	\\u2064
U+2066	Invisible	Very high concern	Tracked, Invisible format, Control	LTR Isolate	Tracked character	-	remove	-	\\u2066
U+2067	Invisible	Very high concern	Tracked, Invisible format, Control	RTL Isolate	Tracked character	-	remove	-	\\u2067
U+2068	Invisible	Very high concern	Tracked, Invisible format, Control	First Strong Isolate	Tracked character	-	remove	-	\\u2068
U+2069	Invisible	Very high concern	Tracked, Invisible format, Control	Pop Directional Isolate	Tracked character	-	remove	-	\\u2069
U+206A	Invisible	Very high concern	Tracked, Invisible format, Control	Inhibit Symmetric Swap	Tracked character	-	remove	-	\\u206a
U+206B	Invisible	Very high concern	Tracked, Invisible format, Control	Activate Symmetric Swap	Tracked character	-	remove	-	\\u206b
U+206C	Invisible	Very high concern	Tracked, Invisible format, Control	Inhibit Arabic Form Shaping	Tracked character	-	remove	-	\\u206c
U+206D	Invisible	Very high concern	Tracked, Invisible format, Control	Activate Arabic Form Shaping	Tracked character	-	remove	-	\\u206d
U+206E	Invisible	Very high concern	Tracked, Invisible format, Control	National Digit Shapes	Tracked character	-	remove	-	\\u206e
U+206F	Invisible	Very high concern	Tracked, Invisible format, Control	Nominal Digit Shapes	Tracked character	-	remove	-	\\u206f
U+20A1	Visible	No concern	Tracked, Safe symbol, Symbol	Colon Sign	Common safe symbol	\\u20a110	keep	-	\\u20a1
U+20A6	Visible	No concern	Tracked, Safe symbol, Symbol	Naira Sign	Common safe symbol	\\u20a610	keep	-	\\u20a6
U+20A9	Visible	No concern	Tracked, Safe symbol, Symbol	Won Sign	Common safe symbol	\\u20a91000	keep	-	\\u20a9
U+20AA	Visible	No concern	Tracked, Safe symbol, Symbol	New Sheqel Sign	Common safe symbol	\\u20aa10	keep	-	\\u20aa
U+20AB	Visible	No concern	Tracked, Safe symbol, Symbol	Dong Sign	Common safe symbol	\\u20ab10	keep	-	\\u20ab
U+20AC	Visible	No concern	Tracked, Safe symbol, Symbol	Euro Sign	Common safe symbol	\\u20ac10	keep	-	\\u20ac
U+20AD	Visible	No concern	Tracked, Safe symbol, Symbol	Kip Sign	Common safe symbol	\\u20ad10	keep	-	\\u20ad
U+20B1	Visible	No concern	Tracked, Safe symbol, Symbol	Peso Sign	Common safe symbol	\\u20b110	keep	-	\\u20b1
U+20B2	Visible	No concern	Tracked, Safe symbol, Symbol	Guarani Sign	Common safe symbol	\\u20b210	keep	-	\\u20b2
U+20B4	Visible	No concern	Tracked, Safe symbol, Symbol	Hryvnia Sign	Common safe symbol	\\u20b410	keep	-	\\u20b4
U+20B9	Visible	No concern	Tracked, Safe symbol, Symbol	Indian Rupee Sign	Common safe symbol	\\u20b9500	keep	-	\\u20b9
U+20BA	Visible	No concern	Tracked, Safe symbol, Symbol	Turkish Lira Sign	Common safe symbol	\\u20ba10	keep	-	\\u20ba
U+20BD	Visible	No concern	Tracked, Safe symbol, Symbol	Ruble Sign	Common safe symbol	\\u20bd100	keep	-	\\u20bd
U+20BF	Visible	No concern	Tracked, Safe symbol, Symbol	Bitcoin Sign	Common safe symbol	\\u20bf1	keep	-	\\u20bf
U+2122	Visible	No concern	Tracked, Safe symbol, Symbol	Trade Mark Sign	Common safe symbol	Name\\u2122	keep	-	\\u2122
U+27E6	Visible	Medium concern	Tracked, Bracket, Punctuation	Mathematical Left White Square Bracket	Bracket alternative	\\u27e6text	[	[	\\u27e6
U+27E7	Visible	Medium concern	Tracked, Bracket, Punctuation	Mathematical Right White Square Bracket	Bracket alternative	text\\u27e7	]	]	\\u27e7
U+27E8	Visible	Medium concern	Tracked, Bracket, Punctuation	Mathematical Left Angle Bracket	Bracket alternative	\\u27e8text	<	<	\\u27e8
U+27E9	Visible	Medium concern	Tracked, Bracket, Punctuation	Mathematical Right Angle Bracket	Bracket alternative	text\\u27e9	>	>	\\u27e9
U+2800	Visible	Medium concern	Tracked, Space, Symbol	Braille Blank	Tracked character	-	? space	-	\\u2800
U+3000	Visible	Medium concern	Tracked, Space	Ideographic Space	Tracked character	-	? space	-	-
U+3010	Visible	Medium concern	Tracked, Bracket, Punctuation	Left Black Lenticular Bracket	Bracket alternative	\\u3010text	[	[	\\u3010
U+3011	Visible	Medium concern	Tracked, Bracket, Punctuation	Right Black Lenticular Bracket	Bracket alternative	text\\u3011	]	]	\\u3011
U+3016	Visible	High concern	Tracked, Bracket, Punctuation	Left White Lenticular Bracket	AI citation bracket alternative	\\u30166\\u2020L9-L11	[	[	\\u3016
U+3017	Visible	High concern	Tracked, Bracket, Punctuation	Right White Lenticular Bracket	AI citation bracket alternative	6\\u2020L9-L11\\u3017	]	]	\\u3017
U+3164	Visible	Medium concern	Tracked, Space, Letter, International char	Hangul Filler	Tracked character	-	? space	-	\\u3164
U+E200	Visible	Very high concern	Tracked, Control	Private Use Citation Start	OpenAI citation marker	\\ue200cite	remove	-	\\ue200
U+E201	Visible	Very high concern	Tracked, Control	Private Use Citation Stop	OpenAI citation marker	\\ue201	remove	-	\\ue201
U+E202	Visible	Very high concern	Tracked, Control	Private Use Citation Delimiter	OpenAI citation marker	\\ue202	remove	-	\\ue202
U+FE00	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-1	Tracked character	-	remove	-	\\ufe00
U+FE01	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-2	Tracked character	-	remove	-	\\ufe01
U+FE02	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-3	Tracked character	-	remove	-	\\ufe02
U+FE03	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-4	Tracked character	-	remove	-	\\ufe03
U+FE04	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-5	Tracked character	-	remove	-	\\ufe04
U+FE05	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-6	Tracked character	-	remove	-	\\ufe05
U+FE06	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-7	Tracked character	-	remove	-	\\ufe06
U+FE07	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-8	Tracked character	-	remove	-	\\ufe07
U+FE08	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-9	Tracked character	-	remove	-	\\ufe08
U+FE09	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-10	Tracked character	-	remove	-	\\ufe09
U+FE0A	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-11	Tracked character	-	remove	-	\\ufe0a
U+FE0B	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-12	Tracked character	-	remove	-	\\ufe0b
U+FE0C	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-13	Tracked character	-	remove	-	\\ufe0c
U+FE0D	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-14	Tracked character	-	remove	-	\\ufe0d
U+FE0E	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-15	Tracked character	-	remove	-	\\ufe0e
U+FE0F	Invisible	Very high concern	Tracked, Invisible format, Mark, International char	Variation Selector-16	Tracked character	-	remove	-	\\ufe0f
U+FEFF	Invisible	Very high concern	Tracked, Invisible format, Control	Zero-Width NBSP / BOM	Tracked character	-	remove	-	\\ufeff
U+FF3B	Visible	Medium concern	Tracked, Bracket, Punctuation	Fullwidth Left Square Bracket	Bracket alternative	\\uff3btext	[	[	\\uff3b
U+FF3D	Visible	Medium concern	Tracked, Bracket, Punctuation	Fullwidth Right Square Bracket	Bracket alternative	text\\uff3d	]	]	\\uff3d
U+FFA0	Visible	Medium concern	Tracked, Space, Letter, International char	Half-width Hangul Filler	Tracked character	-	? space	-	\\uffa0
U+FFFC	Visible	High concern	Tracked, Symbol	Object Replacement	Tracked character	-	? "[OBJECT]"	-	\\ufffc

Holistic AI pattern rules

The holistic rule sections below line up with the section comments in glyphguard.py. The summary tables include the rule metadata used by the app. Regexes are in collapsible blocks after each table because putting the full patterns directly in the table makes the README almost unreadable.

Prose and style heuristics

General prose, style and structure heuristics from the first rule block in glyphguard.py.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
ai_vocabulary	(?i)\b(tapestry|delve|nuance|multifaceted|leverage|holistic|seamless|comprehensive|robust|foster|garner|intricate|poignant|vibrant)\b	medium	1.0	LLM Style, Word Choice	wikipedia_signs_ai_writing	Generic high-polish vocabulary associated with LLM prose. Significance and promotional terms are split into more precise sibling rules.	Prefer concrete wording and remove decorative vocabulary when it adds no meaning.	The report delves into a vibrant tapestry of ideas. The proposal offers a holistic and multifaceted view.
rule_of_three	\b\w{3,},\s+\w{3,},\s+and\s+\w{3,}\b	low	0.5	Structure, List Pattern	wikipedia_signs_ai_writing	Three-item list structure that generated prose can overuse.	Vary list structures or restructure the sentence.	Clearer, faster, and safer delivery matters. Teams need planning, testing, and monitoring.
negative_parallelism	(?i)\b(?:it\s+is|it'?s)\s+not\s+(?:just|merely|only|simply)\s+[^.?!\n]{1,120}?(?:,|;|\u2014|--|\s+but\s+(?:also|rather|instead)\b)	high	2.0	Structure, Rhetorical Pattern	wikipedia_signs_ai_writing	Formulaic not-just contrast framing.	State the contrast directly without the not-just formula.	It's not just speed, it's reliability. It is not merely a tool, but also a platform.
copula_avoidance	(?i)\b(?:serves?\s+as|stands?\s+as|represents?\s+(?:an?\s+)?(?:important|major|significant|key|pivotal|crucial)?\s*\w+|marks?\s+(?:an?\s+)?(?:important|major|significant|key|pivotal|crucial|historic)\s+\w+|holds?\s+the\s+distinction\s+of)\b	medium	1.5	Grammar, Marketing Style, Low Specificity	wikipedia_signs_ai_writing	Ornate replacements for simple is/are descriptions, narrowed to puffery-like copula substitutes.	Prefer direct descriptions where ornate framing adds no evidence.	The gallery serves as the organisation\u2019s exhibition space. The release marks a significant milestone for the project.
prestige_boast_framing	(?i)\b(?:boasts?\s+(?:an?\s+)?(?:impressive|extensive|rich|diverse|robust|unique)?\s*\w+|renowned\s+for|celebrated\s+as|acclaimed\s+for|recognised\s+as|recognized\s+as)\b	high	2.0	Marketing Style, Low Specificity	wikipedia_signs_ai_writing	Prestige or boast framing separated from generic promotional buzzwords.	Replace prestige framing with verifiable facts.	The platform boasts extensive analytics. The artist is renowned for a unique visual language.
promotional_buzzwords	(?i)\b(?:groundbreaking|cutting[-\s]?edge|state[-\s]?of[-\s]?the[-\s]?art|game[-\s]?changer|revolutionary|breathtaking|stunning|unprecedented|unparalleled|transformative|disruptive|innovative\s+(?:solution|approach))\b	high	2.0	Marketing Style, Low Specificity	wikipedia_signs_ai_writing	Standalone promotional buzzwords, separated from copula/boast framing.	Use specific, concrete descriptions instead of buzzwords.	This groundbreaking solution changes the market. The tool offers a state-of-the-art approach.
vague_attribution	(?i)\b(?:industry\s+experts\s+(?:say|believe|agree)|observers\s+(?:note|suggest|point)|critics\s+(?:argue|contend|suggest)|some\s+(?:argue|say|believe|suggest)|many\s+(?:believe|argue|contend|suggest)|it\s+is\s+widely\s+(?:believed|accepted|recognized|recognised|acknowledged)|research\s+(?:shows|suggests|indicates|demonstrates)\s+that)\b	high	2.0	Evidence Gap, Attribution	wikipedia_signs_ai_writing	Vague attribution without a named source or citation.	Cite a specific source or remove the claim.	Industry experts say adoption will grow. Research suggests that teams are more productive.
collaborative_artifact	(?i)\b(?:I\s+hope\s+this\s+helps|let\s+me\s+know\s+if|would\s+you\s+like\s+(?:me\s+)?to|of\s+course!|certainly!|absolutely!|I'?d\s+be\s+(?:happy|glad|pleased)\s+to\s+help|is\s+there\s+anything\s+else|feel\s+free\s+to\s+ask)\b	critical	3.0	Chatbot Artefact	wikipedia_signs_ai_writing	Chatbot conversation artefacts left in final text.	Remove these phrases entirely.	I hope this helps. Let me know if you want changes.
formal_filler_phrase	(?i)\b(?:in\s+order\s+to|due\s+to\s+the\s+fact\s+that|at\s+this\s+point\s+in\s+time|without\s+(?:a\s+)?doubt|needless\s+to\s+say|last\s+but\s+not\s+least)\b	medium	1.0	LLM Style, Filler	wikipedia_signs_ai_writing	Formal filler phrases that add bulk without much information.	Cut the phrase and start directly with the main point.	In order to deploy, run the script. Due to the fact that costs vary, estimates changed.
ing_superficial	\b\w+ing,\s+(?:ensuring|reflecting|highlighting|underscoring|emphasizing|emphasising|demonstrating|illustrating|showcasing)\b	high	2.0	Structure, Rhetorical Pattern	wikipedia_signs_ai_writing	Superficial -ing analysis clause used as a shortcut.	Replace with direct statements about cause, effect or evidence.	The update improves caching, ensuring faster loads. The rollout supports scaling, highlighting adoption.
negative_conclusion	(?i)\b(?:despite\s+(?:these\s+|the\s+|some\s+)?\w*\s*challenges?|faces?\s+several\s+challenges?|moving\s+forward|looking\s+ahead|going\s+forward|as\s+we\s+(?:move|look)\s+(?:forward|ahead))\b	medium	1.0	Essay Style, Structure	wikipedia_signs_ai_writing	Generic challenge/future-outlook framing used to sound balanced.	Be specific about the challenge or future action.	Despite these challenges, the project continued. Looking ahead, the team plans to iterate.
hedging	(?i)\b(?:it\s+(?:seems|appears)\s+(?:that|to\s+be)|arguably|one\s+might\s+(?:say|argue|suggest)|it\s+could\s+be\s+(?:said|argued)\s+that|in\s+(?:many|some)\s+(?:ways|respects)|to\s+(?:some|a\s+(?:certain|large))\s+extent)\b	medium	1.0	Cautious Style, Low Specificity	wikipedia_signs_ai_writing	Hedging language that weakens or genericises claims.	State the claim directly when evidence supports it.	It seems that the model is improving. Arguably, the trade-off is acceptable.
conclusion_transition	(?i)\b(?:in\s+conclusion|to\s+sum\s+up|wrapping\s+up|to\s+(?:recap|recapitulate)|all\s+in\s+all|in\s+(?:the\s+)?final\s+analysis)\b	medium	1.5	Essay Style, Structure, Transition	wikipedia_signs_ai_writing	Formulaic ending transitions. Summary-specific variants are split into section_summary_transition.	End naturally or use a specific summary point.	In conclusion, the change worked. To sum up, the release was successful.
conjunctive_adverb_overuse	(?i)\b(?:furthermore|moreover|additionally|consequently|nevertheless|nonetheless|notwithstanding|whereas|hereby|thereof|therein|wherein)\b	medium	1.5	Formal Style, Transition	wikipedia_signs_ai_writing	Formal conjunctive adverbs that can make prose mechanically academic.	Use simpler transitions where clearer.	Furthermore, the API supports retries. Nevertheless, latency remains a risk.
universal_claim	(?i)\b(?:everyone\s+(?:knows|agrees|understands)|no\s+one\s+(?:can\s+)?deny|it'?s?\s+universally\s+(?:accepted|acknowledged)|we\s+all\s+(?:know|agree|understand)|across\s+the\s+(?:board|globe|world))\b	high	2.0	Evidence Gap, Overclaim	wikipedia_signs_ai_writing	Universal generalisations used to imply consensus without evidence.	Qualify the claim or provide evidence.	Everyone knows this approach is best. No one can deny the impact.
numbered_benefits	(?i)\b(?:three|3|four|4|five|5)\s+(?:key|main|primary|major|significant)\s+(?:benefits|advantages|reasons|factors|ways)\b	medium	1.5	List Pattern, Structure	wikipedia_signs_ai_writing	Listicle-style numbered benefit or reason framing.	Introduce benefits naturally instead of announcing the count.	Three key benefits are speed, safety, and cost. 5 main reasons explain the change.

Citation and tool-output leaks

OpenAI citation, source, prompt and UI leak rules.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
openai_browser_line_citation	(?:\u3010|\u3016)\s*\d+\s*\u2020\s*L\d+(?:-L\d+)?\s*(?:\u3011|\u3017)	critical	5.0	Tool Artefact, Citation Leak, OpenAI, ChatGPT, Harmony	openai_harmony_docs	Raw OpenAI browser, Deep Research or Harmony-style line citation marker.	Render the citation properly or strip the marker.	See the source here \u30166\u2020L9-L11\u3017. The claim appears at \u30109\u2020L9-L11\u3011.
openai_assistants_source_placeholder	(?:\u3010|\u3016)\s*\d+:\d+\s*[\u2020\u2021]\s*[^\u3011\u3017\r\n]{1,120}\s*(?:\u3011|\u3017)	critical	5.0	Tool Artefact, Citation Leak, OpenAI, Assistants, File Search	stackoverflow_openai_assistants_placeholder	Raw OpenAI Assistants/File Search placeholder citation.	Replace with a visible citation or strip it.	Raw marker \u30164:2\u2020source\u3017 appeared. Another marker was \u301012:1\u2021source\u3011.
openai_private_use_citation	\uE200cite\uE202(?:[A-Za-z][\w.-]*|turn(?:\d+|X)\w*(?:\d+|Y)?)(?:\uE202(?:L\d+(?:-L\d+)?|P(?:aragraph)?\s*\d+|[A-Za-z][\w.-]*|turn\d+\w+\d+))*\uE201	critical	5.0	Tool Artefact, Citation Leak, OpenAI, API	openai_citation_formatting_docs, stackoverflow_openai_assistants_placeholder	Raw OpenAI private-use Unicode citation marker.	Remove or render the citation properly.	\uE200cite\uE202turn0file1\uE202L8-L13\uE201 \uE200cite\uE202turnXfileY\uE201
openai_private_use_generic_marker	\ue200(?!cite\ue202)[^\r\n\ue201]{1,240}\ue202[^\r\n\ue201]{1,240}\ue201	critical	4.5	Tool Artefact, Citation Leak, OpenAI, ChatGPT	openai_citation_formatting_docs, observed_chatgpt_content_reference	Raw OpenAI/ChatGPT private-use UI marker with an internal delimiter.	Remove or render the marker properly.	\uE200filecite\uE202turn0file0\uE201 \uE200products\uE202{"id":1}\uE201
openai_content_reference_marker	:contentReference\[(?:oaicite|[^\]\s:]+):\d+\]\{index=\d+\}	critical	4.0	Tool Artefact, Citation Leak, OpenAI, ChatGPT	wikipedia_signs_ai_writing, observed_chatgpt_content_reference	Raw ChatGPT/OpenAI contentReference citation marker.	Remove or render the reference properly.	:contentReference[oaicite:3]{index=3} :contentReference[source:17]{index=17}
openai_naked_source_id	(?<!\ue202)\bturn\d+(?:file|search|news|url|block|image|product|finance|forecast|sports|time)\d+\b	high	3.0	Tool Artefact, Citation Leak, OpenAI, API	openai_citation_formatting_docs, wikipedia_signs_ai_writing	OpenAI-style source reference ID leaked without its citation wrapper.	Check whether a raw citation/source ID leaked.	turn0file0 turn2search5
openai_harmony_special_token	<\|(?:start|end|message|channel|constrain|call|return)\|>	critical	5.0	Tool Artefact, Prompt Leak, OpenAI, Harmony	openai_harmony_docs	OpenAI Harmony prompt/message special token leaked into visible text.	Remove internal prompt-format tokens.	<|start|> <|message|>
openai_tool_channel_leak	\bValid channels:\s*analysis\s*,\s*(?:commentary\s*,\s*)?final\b|\bto=(?:functions|browser|python|web|file_search|image_gen|automations)\.[A-Za-z_]\w*\b	critical	4.0	Tool Artefact, Prompt Leak, OpenAI, Tool use	openai_harmony_docs	OpenAI-style tool/channel instruction leaked into visible text.	Remove internal tool-call or channel-routing text.	Valid channels: analysis, commentary, final to=functions.get_current_weather

Metadata and API leaks

Cross-vendor API and grounding metadata rules.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
openai_file_citation_json	"type"\s*:\s*"file_citation"|"annotations"\s*:\s*\[|"file_id"\s*:\s*"file-[A-Za-z0-9_-]+	high	3.0	Metadata Leak, API Leak, OpenAI, File Search	openai_file_search_docs	Raw OpenAI Responses/File Search citation annotation JSON.	Treat as API/debug output rather than final prose.	"type": "file_citation" "file_id": "file-abc123"
anthropic_citation_json	"cited_text"\s*:|"document_index"\s*:|"start_char_index"\s*:|"end_char_index"\s*:|"start_page_number"\s*:|"end_page_number"\s*:|"start_block_index"\s*:|"end_block_index"\s*:	high	2.5	Metadata Leak, API Leak, Anthropic, Claude	anthropic_claude_citations_docs	Raw Anthropic/Claude citation metadata.	Treat as API metadata rather than final prose.	"cited_text": "..." "start_char_index": 42
gemini_grounding_metadata_json	"groundingMetadata"\s*:|"webSearchQueries"\s*:|"groundingChunks"\s*:|"groundingSupports"\s*:|"groundingChunkIndices"\s*:|"searchEntryPoint"\s*:	high	2.5	Metadata Leak, API Leak, Google, Gemini	google_gemini_grounding_docs	Raw Gemini grounding/citation metadata.	Treat as Gemini API grounding metadata.	"groundingMetadata": {} "webSearchQueries": ["x"]
perplexity_search_metadata_json	"citations"\s*:\s*\[|"search_results"\s*:\s*\[|"search_context_size"\s*:	medium	2.0	Metadata Leak, API Leak, Perplexity, Sonar	perplexity_sonar_features_docs	Raw Perplexity/Sonar search or citation metadata.	Treat as API/search metadata rather than final prose.	"citations": ["https://example.com"] "search_context_size": "low"

Model/vendor-specific artefacts

Model/vendor-specific rules where the marker is narrow enough to tag.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
openai_creature_metaphor_tic	(?i)\b(?:(?:little|tiny|small|feral|chaos|formatting|file[-\s]?format|filesystem|spreadsheet|cache|parser|prompt|token|tool|build|deploy(?:ment)?|perf|regex|yaml|json|state[-\s]?machine|clipboard|merge|git|branch|tooltip|ui|ux)\s+(?:goblin|goblins|gremlin|gremlins|troll|trolls|ogre|ogres|raccoon|raccoons|pigeon|pigeons)|(?:goblin|gremlin|troll|ogre|raccoon|pigeon)[-\s]?(?:mode|math|brain|energy|gearbox|bandwidth|version|logic))\b	high	3.0	LLM Style, Word Choice, LLM Metaphor, OpenAI, Codex, GPT-5.1+, GPT-5.4-Nerdy, GPT-5.5-Codex	openai_goblin_article, tidbits_gpt55_prompt_discussion	Creature-metaphor lexical tic associated with OpenAI GPT-5.1+ / Codex style leakage.	Review whether the metaphor is intentional.	The filesystem goblin ate the config. There is a little perf gremlin running unattended.

Dash-delimited parenthetical asides

Dash-aside punctuation heuristics kept separate because they are noisy unless repeated.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
unicode_dash_delimited_parenthetical	(?<!\S)(?:\u2014\s*[^\n\u2014]{8,180}\s*\u2014|\u2013\s*[^\n\u2013]{8,180}\s*\u2013)(?!\S)	medium	2.0	Punctuation Pattern, Parenthetical Aside, Structure	wikipedia_signs_ai_writing	Unicode dash-delimited parenthetical aside using paired em/en dashes.	Consider commas, parentheses or a separate sentence.	The parser \u2014 somehow still? alive \u2014 handled the payload. The model \u2013 trying to sound casual \u2013 added another aside.
ascii_dash_delimited_parenthetical	(?<!\S)(?:--\s*(?=[^\n]{8,180}?\s*--)(?=[^\n]*[A-Za-z])(?:(?!--)[^\n]){8,180}?\s*--|-\s+(?=[^\n]{8,180}?\s+-)(?=[^\n]*[A-Za-z])(?:(?!\s+-)[^\n]){8,180}?\s+-)(?!\S)	low	1.25	Punctuation Pattern, Parenthetical Aside, Structure	wikipedia_signs_ai_writing	ASCII dash-delimited parenthetical aside.	Weak structural signal unless it appears with other markers.	The cache -- because, naturally, it hates us -- lied. The UI - in its infinite wisdom - ate the tooltip.

Wikipedia lead, prose and formatting heuristics

Rules for Wikipedia-style lead, prose and formatting patterns.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
leads_list_title_proper_noun	(?im)^\s*(?:The\s+)?(?:[\u201C"]?List\s+of\s+[^\n.]{3,120}[\u201D"]?|[A-Z][A-Za-z0-9\u2019\'-]+(?:\s+[A-Z][A-Za-z0-9\u2019\'-]+){0,4}\s+editions)\s+(?:is|are|refers\s+to|constitutes|comprises)\b	medium	1.0	Structure, Lead	wikipedia_signs_ai_writing	Lead sentence treats a list/article title as a standalone proper noun.	Introduce lists and descriptive titles naturally.	The \u201CList of songs about Mexico\u201D is a curated compilation of musical works. EuroGames editions is the chronological list of the biennial EuroGames.
negative_parallelism_not_only	(?i)\bnot\s+only\b(?:[^.?!\n]{0,180}\bbut\s+(?:also|rather|instead)\b)?	medium	1.5	Structure, Rhetorical Pattern	wikipedia_signs_ai_writing	Broad not only / but also parallelism marker.	Use direct comparison or contrast where possible.	This choice is not only dismissive but also harsh. The portrait constitutes not only self-representation but a visual document.
negative_parallelism_not_but	(?i)\b(?:it\s+is|it'?s|this\s+is|that\s+is)?\s*not\s+(?:an?\s+)?[^.?!\n\u2014-]{1,90}?\s+(?:but|rather)\b|\bnot\s+[^.?!\n\u2014-]{1,40}?,\s*not\s+[^.?!\n\u2014-]{1,40}?,\s*not\s+[^.?!\n\u2014-]{1,80}?(?:\u2014|--|,)?\s*(?:just|only|rather|but)\b	high	2.0	Structure, Rhetorical Pattern	wikipedia_signs_ai_writing	Separate negative-parallelism shape for not X but Y and repeated not-list contrast.	Avoid formulaic rejection/reframe statements.	It is not a career but an algorithmic moment. Not a career, not a body of work, not sustained relevance \u2014 just an algorithmic moment.
title_case_headings	(?m)^\s*(?:#{1,6}\s*)?(?=[^\n]{12,120}$)(?:[A-Z][A-Za-z0-9\u2019\'-]+|of|and|the|for|in|on|to|with)(?:\s+(?:[A-Z][A-Za-z0-9\u2019\'-]+|of|and|the|for|in|on|to|with)){2,}\s*$	none	0.25	No Concern, Headings, Formatting	wikipedia_signs_ai_writing	Title Case heading-like line. Added as an opt-in/no-concern review signal because this is noisy outside heading context.	Only flag if the user opts into no-concern/low-confidence style review.	Impact of Technology and Digitalization Sustainable Development and Environmental Law
boldface_inline_header_list	(?m)^\s*(?:[-*\u2022]|\d+[.)])\s+(?:(?:'''[^'\n:]{2,80}:?'''|\*\*[^*\n:]{2,80}:?\*\*)|(?:'''[^'\n]{2,80}'''|\*\*[^*\n]{2,80}\*\*)\s*:)\s+	medium	1.5	Structure, List Pattern, Formatting	wikipedia_signs_ai_writing	List item begins with a bold inline header and colon.	Use proper list markup and avoid slide-deck style inline headers.	1. **Historical context:** The world was changing after WWII. - **Conflict of interest:** Please review the sourcing.
boldface_repeated_spans	(?s)(?:'''[^'\n]{2,120}'''|\*\*[^*\n]{2,120}\*\*)(?:(?!\n\s*\n).){0,300}(?:'''[^'\n]{2,120}'''|\*\*[^*\n]{2,120}\*\*)	low	1.0	Formatting, Low Specificity	wikipedia_signs_ai_writing	Multiple bold spans inside one paragraph or compact block.	Use emphasis sparingly.	**Background** text continues, then **Impact** appears in the same paragraph. '''History''' is followed by prose and '''Reception''' in one paragraph.
inline_header_list	(?m)^\s*(?:[-*\u2022#]|\d+[.)])\s+(?!'''|\*\*)(?:[A-Z][^:\n]{2,80})\s*:\s+	medium	1.0	Structure, List Pattern	wikipedia_signs_ai_writing	List item with a colon-headed mini-section.	Integrate short items into prose or use proper headings/lists.	\u2022 Historical context: The world was changing after WWII. 1. Historical context post-WWII era: The world was rapidly changing.
thematic_break_heading	(?m)^\s*----\s*\n\s*==+\s*[^\n=]+\s*==+	medium	1.0	Structure, Markup, Wikitext	wikipedia_signs_ai_writing	Markdown-like thematic break immediately before a wikitext heading.	Remove extraneous horizontal rules before headings.	----<br>== History == ----<br>=== Form and construction ===

Chatbot disclaimers, placeholders and renderer artefacts

Rules for copied chatbot disclaimers, placeholder text, renderer leaks and tracking parameters.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
knowledge_cutoff_disclaimer	(?i)\b(?:as\s+of\s+(?:my\s+)?(?:last\s+)?(?:knowledge|training)\s+update|up\s+to\s+my\s+last\s+training\s+update|my\s+knowledge\s+(?:only\s+)?(?:goes|extends)\s+(?:up\s+)?to)\b	critical	2.5	Chatbot Artefact, Evidence Gap, Disclaimer	wikipedia_signs_ai_writing	Explicit model knowledge/training cutoff disclaimer.	Remove chatbot-cutoff text and verify current sources.	As of my last knowledge update in January 2022, I do not have current information. Up to my last training update, this was not listed.
source_scarcity_disclaimer	(?i)\b(?:while\s+specific\s+(?:details|information)\s+(?:is|are)\s+(?:limited|scarce)|not\s+widely\s+(?:available|documented|disclosed)|based\s+on\s+available\s+information|maintains\s+a\s+low\s+profile|keeps\s+personal\s+details\s+private)\b	medium	1.5	Evidence Gap, Source Scarcity, Disclaimer	wikipedia_signs_ai_writing	Source-scarcity speculation often used to pad missing evidence.	State only documented information or mark gaps explicitly.	Specific details are not widely documented in available sources. Based on available information, the village has limited coverage.
placeholder_text	(?i)(?:\[\s*(?:Describe|Enter|Your Name|specify|Provide)[^\]\n]{0,120}\]|\b(?:INSERT|PASTE|SOURCE|URL|ACCESS[_-]DATE|DATE)_[A-Z0-9_]+\b|\b\d{4}-(?:XX|xx)-(?:XX|xx)\b)	critical	2.5	Placeholder, Template, Chatbot Artefact	wikipedia_signs_ai_writing	Template placeholder or fill-in-the-blank token left in final text.	Replace placeholders with real content or remove them.	[Describe the specific section that needs editing] SOURCE_URL_30 and access-date=2025-XX-XX remain in the citation.
openai_oai_citation_renderer_leak	(?i)(?:(?<!:)contentReference\[oaicite:\d+\](?!\{index=\d+\})|oai_citation|attached_file:\d+|Example\+\d+)	critical	4.0	Tool Artefact, Citation Leak, OpenAI, UI marker	observed_chatgpt_content_reference, wikipedia_signs_ai_writing	OpenAI-ish citation renderer or attachment marker leak not covered by the stricter contentReference rule.	Remove renderer artefacts or replace with proper citations.	contentReference[oaicite:3] attached_file:2 was left in the copied answer.
grok_citation_renderer_leak	(?i)\b(?:grok_card|grok_render_citation_card_json|referrer=grok\.com)\b	high	3.5	Tool Artefact, Citation Leak, Grok, UI marker	wikipedia_signs_ai_writing	Grok/X citation renderer artefact.	Strip renderer metadata and add normal citations.	grok_card appeared in the copied text. grok_render_citation_card_json was pasted into the article.
triple_colon_writing_block	(?m)^:::[^\W\d_][\w-]*\{[^}\n]*(?:variant|variante|id)\s*=	critical	4.0	Tool Artefact, UI marker, Chatbot Artefact	observed_chatgpt_content_reference	ChatGPT writing-block fence leaked into text, including non-English block names.	Remove the UI wrapper and keep only the intended content.	:::writing{variant="document" id="12345"} :::\u00E9criture{variante="document" id="67890"}
ai_referral_url_parameter	(?i)\b(?:utm_source|utm_medium|utm_campaign|referrer)=(?:openai|chatgpt(?:\.com)?|copilot(?:\.microsoft\.com|\.com)?|grok(?:\.com)?|perplexity|claude|anthropic|gemini|bard)\b	high	3.5	URL Parameter, Citation Leak, Tool Artefact	wikipedia_signs_ai_writing	AI-tool referral/tracking parameter copied with a URL.	Strip tracking parameters and verify the source URL.	https://example.com/?utm_source=chatgpt.com https://x.example/?referrer=grok.com
generic_utm_parameter	(?i)\b(?:utm_source|utm_medium|utm_campaign|utm_term|utm_content)=([^\s&#]+)	low	0.5	URL Parameter, Tracking Parameter	wikipedia_signs_ai_writing	Generic UTM tracking parameter. Not AI-specific, but useful citation hygiene.	Remove tracking parameters when preserving citations.	https://example.com/?utm_source=newsletter https://site.test/?utm_campaign=spring

Wikipedia talk-page, AfC and edit-summary artefacts

Rules for talk-page phrasing, subject-line comments and edit-summary style artefacts.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
canned_quality_goodfaith	(?i)\b(?:aligns?\s+with\s+Wikipedia(?:['\u2019]s)?\s+(?:aim|goal|mission|standards)|adheres?\s+to\s+(?:good\s+faith|neutrality|Wikipedia(?:['\u2019]s)?\s+principles)|assume\s+good\s+faith|constructive\s+and\s+collaborative)\b	low	0.75	Talk Page, Good Faith, Formal Style	wikipedia_signs_ai_writing	Canned talk-page good-faith or quality language.	Use only where it is genuinely part of the discussion.	This aligns with Wikipedia\u2019s mission. Please assume good faith and keep the discussion constructive.
canned_constructive_criticism	(?i)\b(?:if\s+you\s+have\s+any\s+(?:concerns|suggestions|feedback)|if\s+there\s+are\s+any\s+issues|constructive\s+(?:criticism|feedback)|please\s+feel\s+free\s+to\s+(?:improve|comment|respond))\b	low	0.5	Talk Page, Constructive Criticism, Chatbot Artefact	wikipedia_signs_ai_writing	Generic constructive-feedback boilerplate.	Keep only if it adds context to the conversation.	If you have any concerns, please let me know. Constructive feedback is welcome.
canned_focus_on_content	(?i)\b(?:let\s*'?s\s+focus\s+on\s+(?:improving\s+)?(?:the\s+)?(?:article|content)|in\s+the\s+spirit\s+of\s+collaboration|focus\s+on\s+content\s+rather\s+than\s+contributors)\b	low	0.5	Talk Page, Focus Shift	wikipedia_signs_ai_writing	Canned talk-page focus-shift wording.	Use specific article/content issues instead of generic moderation language.	Let's focus on improving the article. Please focus on content rather than contributors.
subject_line_comment	(?m)^\s*Subject:\s+[^\n]{1,160}$	low	0.75	Talk Page, Subject Line, Template	wikipedia_signs_ai_writing	Email/form-style subject line left inside a comment.	Remove the subject line or convert it into a heading only if appropriate.	Subject: Request for review Subject: Sources for this article

Wikitext markup, citation-quality and category artefacts

Rules for small tables, headings, categories, reference reuse, book citation pages and policy shortcut clusters.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
unusual_small_table	(?s:\{\x7c(?=.{0,700}\x7c\})(?=(?:[^\n]*\n){1,10})[^{}]{0,700}\x7c\})|(?m:^\s*\x7c\s*(?:Path|Key|Source|Item|Name)\s*\x7c[^\n]*\x7c\s*\n\s*\x7c(?:\s*:?-{3,}:?\s*\x7c){2,5}\s*\n(?!(?:\s*\x7c[^\n]*\x7c\s*\n?){7})(?:\s*\x7c[^\n]*(?:\[[^\]]+\]\([^)]+\)|https?://)[^\n]*\x7c\s*\n?){1,6})	low	0.75	Table, Markup, Wikitext	wikipedia_signs_ai_writing	Very small wikitext table, or small generated-looking Markdown index/link table.	Review whether the table is complete and valid. For short link/index lists, use bullets unless a table adds real structure.	{| class="wikitable"<br>| Example<br>|} {|<br>! A<br>| B<br>|} | Path | Description |<br>| --- | --- |<br>| [samples/x](samples/x) | Example sample set. |
skip_heading_levels	(?m)^\s*={3,}\s*[^=\n][^\n]*={3,}\s*$	low	0.5	Headings, Markup, Wikitext	wikipedia_signs_ai_writing	Potential skipped heading level in wikitext.	Check heading hierarchy.	=== History === ==== Reception ====
broken_category_syntax	\[\[Category:[^\]\|\n]{0,120}[/<>][^\]\n]{0,120}\]\]	high	2.0	Category, Markup, Malformed	wikipedia_signs_ai_writing	Malformed category link with slash or angle-bracket placeholder.	Fix or remove malformed category syntax.	[[Category:Example/Bad]] [[Category:<Insert category>]]
attributable_index_json	(?:\(attribution\)\{[^}\n]*"attributableIndex"|"attributableIndex"\s*:\s*"?\d+(?:-\d+)?"?)	high	2.5	Metadata Leak, API Leak, Citation Leak	wikipedia_signs_ai_writing	Attribution/citation JSON metadata leaked into text.	Strip API metadata and retain only visible citations.	(attribution){"attributableIndex":"1-2"} "attributableIndex": "7"
invalid_reference_reuse	(?is)<ref\b[^>]*>[^<]{0,400}<ref\b|</ref>\s*</ref>	medium	1.5	Markup, Citation Leak, Malformed	wikipedia_signs_ai_writing	Nested reference tag or duplicate closing reference tag.	Fix the reference markup before publication.	<ref>first <ref>nested</ref> citation</ref></ref>
book_citation_missing_page	(?is)\{\{\s*cite\s+book\b(?=[^}]{0,600}\|\s*title\s*=)(?![^}]{0,600}\|\s*pages?\s*=)[^}]{0,600}\}\}	none	0.25	No Concern, Citation Quality, Book Citation	wikipedia_signs_ai_writing	Book citation without page/pages field. Citation-quality signal, not AI evidence.	Only flag in a future citation-quality/no-concern mode.	{{cite book |title=Example book |author=A. Writer}} {{ cite book | title=Long book | publisher=Press }}
edit_summary_first_person	(?i)\b(?:I\s+(?:revised|updated|improved|adjusted|expanded|refined|corrected|added|removed|ensured|avoided)\b|my\s+(?:edit|changes|summary)\b)	low	0.75	Edit Summary, First Person	wikipedia_signs_ai_writing	First-person edit-summary language in copied text.	Check whether an edit summary or chatbot response was pasted into content.	I revised the article for neutrality. My edit improves the sourcing.
afc_submission_statement	(?i)(?:\bReviewer\s+note\s*\(for\s+AfC\)|\bThis\s+draft\s+is\s+a\s+neutral\s+and\s+well[-\s]sourced\s+biography\b|\bAfC\s+submission\b)	high	2.0	AfC, Template, Markup	wikipedia_signs_ai_writing	AfC-specific generated/review artefact.	Remove review boilerplate from article prose.	Reviewer note (for AfC): this draft has been improved. This draft is a neutral and well-sourced biography.
wikilawyering_multiple_shortcuts	(?im)^(?=(?:.*\bWP:[A-Z0-9]{1,}\b){3,}).{0,400}$	medium	1.25	Talk Page, Policy	wikipedia_signs_ai_writing	Talk-page line with multiple policy shortcuts. Useful but not necessarily AI.	Review whether the policy references are meaningful and relevant.	WP:NPOV WP:V WP:OR all apply here. This violates WP:RS, WP:BLP and WP:UNDUE.

Emoji, legacy, incomplete-copy and notability heuristics

Lower-confidence formatting, notability, summary and incomplete-copy rules.

Rule	Regex	Severity	Weight	Tags	Sources	Description	Suggestion	Examples
emoji_bulletpoint	(?m)^\s*[\U0001F300-\U0001FAFF\u2600-\u27BF]\s+	low	0.75	Emoji, Emoji Bullet, Formatting	wikipedia_signs_ai_writing	Emoji used as a bullet marker.	Use plain list markup if this is formal text.	\u2705 Completed the rewrite. \u1F539 Added source details.
emoji_formatting_inline	(?m)(?<!^\s)[\U0001F300-\U0001FAFF\u2600-\u27BF]	low	0.5	Emoji, Formatting	wikipedia_signs_ai_writing	Emoji used as inline formatting or emphasis.	Remove decorative emoji from formal content unless intentional.	This looks polished \u2705 Please review the section \u1F50D.
didactic_disclaimer	(?i)\b(?:it\s+is\s+(?:important|worth|essential)\s+to\s+note\s+that|it\s+should\s+be\s+noted\s+that)\b	medium	1.0	LLM Style, Filler	wikipedia_signs_ai_writing	Didactic note-signposting phrase split from generic filler.	Remove the phrase and state the point directly.	It is important to note that costs vary. It should be noted that the source is incomplete.
section_summary_transition	(?i)\b(?:in\s+summary|to\s+summari[sz]e|overall)\b	medium	1.0	Essay Style, Transition	wikipedia_signs_ai_writing	Section-summary transition split from conclusion-transition phrasing.	Use a specific summary point or omit the transition.	In summary, the change worked. Overall, the release was successful.
ai_prompt_refusal	(?i)\b(?:as\s+an\s+AI\s+language\s+model|I\s+cannot\s+(?:assist|help)\s+with|I\s+can\s+not\s+(?:assist|help)\s+with|I\s+am\s+unable\s+to\s+comply)\b	critical	3.0	Chatbot Artefact, Refusal, Prompt Leak	wikipedia_signs_ai_writing	Raw chatbot refusal or model-identity phrase.	Remove prompt-response artefacts from final text.	As an AI language model, I cannot assist with that. I am unable to comply with this request.
abrupt_cutoff	(?m)\b[A-Za-z]{3,}(?:\.{2,}|\u2026)\s*$	low	0.5	Incomplete	wikipedia_signs_ai_writing	Abrupt trailing ellipsis or incomplete copy at end of line.	Check whether text was truncated during copy/paste or generation.	The article continues with additional... The sentence trails off\u2026
significance_legacy_trend	(?i)\b(?:plays\s+a\s+(?:significant|pivotal|crucial)\s+role|represents\s+a\s+(?:key|significant|major|pivotal)\s+(?:shift|moment)|marks\s+a\s+(?:turning\s+point|milestone)|serves\s+as\s+a\s+(?:catalyst|focal\s+point)|drives\s+the\s+(?:growth|evolution)\s+of|mirrors\s+(?:broader|wider)\s+trends|embodies\s+the\s+(?:ongoing|contemporary)\s+discourse|underscores\s+the\s+(?:importance|legacy|relevance))\b	medium	1.75	Marketing Style, Legacy Hype, Low Specificity	wikipedia_signs_ai_writing	Phrase-level significance/legacy/trend hype split out from vocabulary and promotional rules.	Replace broad significance claims with concrete sourced facts.	The album plays a significant role in the scene. The release mirrors broader trends in digital media.
canned_notability_media	(?i)\b(?:independent\s+(?:local|regional|national|international)\s+(?:media|news|press)\s+(?:outlets|coverage)|profiled\s+in\s+(?:[^,\n]+,\s*){1,3}(?:and\s+)?[^,\n]+|featured\s+in\s+(?:[^,\n]+,\s*){1,3}(?:and\s+)?[^,\n]+|(?:active|strong)\s+(?:social\s+media|online)\s+presence|on\s+(?:platforms|social\s+media)\s+such\s+as\s+[^.\n]{3,80})\b	medium	1.25	Notability, Evidence Gap, Marketing Style	wikipedia_signs_ai_writing	Canned notability/media-coverage argument pattern.	Name specific reliable independent sources and avoid generic notability claims.	The subject has independent local media coverage. She was profiled in Example News, Local Times, and Arts Weekly.
exhaustive_edit_summary	(?i)\bI\s+(?:revised|updated|improved|ensured|expanded|removed)\s+[^\n]{50,180}\b(?:WP:[A-Za-z]{2,}|Neutral\s+Point\s+of\s+View|Verifiability|Original\s+Research|NPOV|reliable\s+sources)\b	medium	1.5	Edit Summary, First Person, Policy	wikipedia_signs_ai_writing	Verbose first-person edit summary with policy-heavy justification.	Keep edit summaries out of article prose.	I revised the lead, improved sourcing, removed unsourced claims, and ensured compliance with WP:NPOV. I updated the draft to improve sourcing, remove unsupported claims, clarify the lead, and improve Verifiability.
hallucinated_category_name	\[\[Category:[^\]\n]{30,120}\]\]	low	0.5	Category, Hallucination	wikipedia_signs_ai_writing	Improbably long category name. Weak signal; can be legitimate.	Verify that the category exists and is appropriate.	[[Category:Contemporary interdisciplinary artists from the upper valley region]] [[Category:People associated with regional innovation and digital outreach]]

Samples and corpus

Included sample sets

This covers the sample material in the repository, including the public sample corpus and the smaller test fixture sets.

• samples/wikipedia_signs_ai_writing/ - Bundled Wikipedia "Signs of AI writing" example corpus used for detector tuning and scorer checks. Contains 198 .txt samples: 175 quote examples and 23 preformatted examples.
• tests/samples/ - Small text fixtures used by tests and manual checks. Contains 6 direct .txt files, including short/long AI-style burger essays, a medium article sample, OpenAI Harmony page text and quick smoke-test text.
• tests/samples/ai_filetype_detection_samples/ - File-type extraction fixture set used by tests/test_score_samples.py. Contains 36 files covering plain text, code, CSV/TSV, JSON/JSONL, YAML, TOML, INI, properties, log, Markdown, LaTeX, HTML/XML, EML, EPUB, XLSX, DOCX, ODT, RTF, PDF, image, binary and SQLite cases.

The Wikipedia corpus filenames include the index, block type and source section slug, for example:

001_quote_undue_emphasis_on_significance_legacy_and_broader_trends.txt
066_pre_not_x_but_y.txt

The file-type fixture names are deliberately literal, for example:

sample_markdown_fenced_block.md
sample_word_document.docx
sample_searchable_text_pdf.pdf

Generated score reports

score_samples.py writes these files by default under results/ or another folder passed with --out:

• file_scores.csv - one row per scored file, including source kind, match counts and scores.
• rules_by_file.csv - one row per matched rule per file.
• rule_frequency.csv - aggregate rule frequency across the scanned files.
• skipped_files.csv - files ignored or failed during extraction.
• file_scores.json - full structured output.

The scorer imports local glyphguard.py headlessly and uses the same rule loader, text extractor and score-breakdown methods as the app.

Adding new samples

Add new corpora under samples/<corpus_name>/ as plain .txt files. Keep one source block per file where possible. Name files with an index, block type and short source slug so sorting is stable and the source is still recognisable.

If the corpus was generated from a page or document, include a manifest with source path, block kind, character count and word count so future scoring runs are traceable. Then run the scorer and review the generated CSV/JSON output before committing sample changes.

Example:

python score_samples.py samples/<corpus_name> --out results

Testing

GlyphGuard has unit tests for the app/rule behaviours, a one-test-per-rule suite for holistic rules, and scorer tests.

Run tests

Run the full test suite:

python -m pytest -q

Run the main app/rule behaviour tests:

python -m pytest -q tests/test_glyphguard.py

Run the one-test-per-rule suite:

python -m pytest -q tests/test_holistic_rules.py

Run scorer tests:

python -m pytest -q tests/test_score_samples.py

Rule test structure

tests/test_holistic_rules.py tests every production pattern explicitly. It includes flattened examples, extra positive and negative cases, metadata checks and overlap guards where similar rules could double-count the same text.

Sample smoke tests

python score_samples.py samples/wikipedia_signs_ai_writing --out /tmp/glyphguard_score_smoke
python score_samples.py samples/wikipedia_signs_ai_writing --limit 10 --out /tmp/glyphguard_score_smoke_limited

Project structure

This repository contains the following files and folders:

• glyphguard.py - Main application module with UI, text extraction, detection logic, scoring, cleaning and report export.
• README.md - Project overview and usage notes.
• score_samples.py - Headless scorer for files and folders.
• samples/ - Bundled corpora used for tuning and regression checks.
• samples/wikipedia_signs_ai_writing/ - Extracted Wikipedia example-block corpus.
• tests/test_glyphguard.py - Unit tests for core app and rule behaviours.
• tests/test_holistic_rules.py - One-test-per-rule suite for the holistic rules.
• tests/test_score_samples.py - Tests for the headless scorer.
• docs/ - Screenshots, generated UI images and sample report pages.
• docs/generate_screenshots.py - Helper for regenerating documentation screenshots.
• CHANGELOG.md - User-visible change history.
• LICENSE - MIT licence text.

Configuration

Preferences are saved to glyphguard_settings.json. The app first tries to write that file beside glyphguard.py; if that is not writable it falls back to ~/.glyphguard_settings.json.

Saved settings include:

• per-character detection toggles;
• replacement toggles and replacement overrides;
• detection-table visibility and category filters;
• panel collapsed/expanded state and remembered widths.

Delete the settings file to reset preferences.

The main rule data lives in these constants:

• TRACKED_CHAR_SPECS - glyph-level detections and replacement defaults.
• HOLISTIC_AI_PATTERN_SPECS - holistic writing-pattern and tool-artefact rules.
• HOLISTIC_SOURCE_CATALOG - source metadata used by holistic rules.
• HOLISTIC_TAG_COLOURS - colour mapping for semantic tags.

Reports and output files

The UI can export:

• PDF reports with summary scores, charts, glyph rows, holistic rule rows, original text and cleaned text.
• Plain-text reports with the same review information in a simpler form.

The scorer writes CSV and JSON output under results/ by default, or the folder passed with --out.

Limitations

• Results are heuristic and can be wrong.
• The PDF extractor is small and stdlib-only. It works best on text-based PDFs and does not do OCR.
• DOCX, ODT and RTF extraction is best-effort visible-text extraction, not a full document renderer.
• It does not inspect images, screenshots, white-on-white text, CSS tricks or layout-level hiding.
• Some legitimate writing will hit AI-style rules. That is expected.
• A clean scan does not prove a document is safe, human-written or free from hidden intent.

Ideas/Roadmap

Potential improvements:

• American/British/International English/Multilingual modes
  • american mode picks up british stuff
  • British/International English modes pick up american spelling and stuff
  • British takes Welsh letters as No concern
  • European discounts all European native languge characters (broad).
  • Multilingual ignores all region based characters as No concern.
• rework report completely, utter garbage atm
• cleaner command-line output for batch scoring
• better generated screenshot flow
• richer sample corpus metadata
  • need actual docs of different types with real examples
  • need to find a good source of tagged AI stuff
  • need to find a good detector of AI slop
• Webpage detector/scorer - Help avoid AI slop webpages quickly
  • browser plugin, colud do taperscript?
  • wonder how many pages have hidden chars (wonder if its a SEO way to get more traffic from LLMS lol)
• stronger extraction for awkward PDFs and document formats
• more file support
• AI code detection speific rules/mode (gotta look up research papers about this)
• more model/vendor specific detections.
• model/vendor guess when evidence is statistically strong enough.
• optional packaging or a requirements file

Contributing

Contributions are welcome. Try to keep the app local and practical.

Holistic rules need to have sources or strong evidence of some kind, please reference your sources!

Please include tests and documentation updates when changing user-visible behaviour, adding rules etc.

Avoid uncommon heavyweight dependencies unless they are clearly justified.

Adding a new glyph rule

Add an entry to TRACKED_CHAR_SPECS with the code point, visibility, description, example, replacement behaviour, keyboard equivalent, HTML/raw value and any kind or concern override. Then add or update tests in tests/test_glyphguard.py for detection, concern level and cleaning behaviour.

If the character is not explicitly tracked, check whether the derived Unicode metadata path already handles it. Add a curated entry when the character has a specific replacement, source, concern level or explanation that the fallback metadata cannot know.

Adding a new holistic rule

Add a rule to HOLISTIC_AI_PATTERN_SPECS. Include:

• name
• regex
• weight
• severity
• examples
• description
• suggestion
• sources
• tags

Use a stable snake_case name. Keep the regex narrow enough to explain. Add examples that show the intended match, and add negative tests for wording that should not match. If you use a new source key, add it to HOLISTIC_SOURCE_CATALOG. If you use a new tag, add it to HOLISTIC_TAG_COLOURS. Add an explicit test in tests/test_holistic_rules.py, including overlap guards where nearby rules might also match.

Adding or updating samples

Keep one source block per .txt file. Use clear filenames with an index, block type and source slug. If possible, add a manifest. Then run:

python score_samples.py samples/<corpus_name> --out results

Review file_scores.csv, rules_by_file.csv, rule_frequency.csv and file_scores.json before committing. If a rule change deliberately shifts scores across the corpus, mention that in the changelog.

Updating tests

Update tests when changing glyph rules, holistic rules, sample extraction, score output or overlap behaviour. Rule changes should include positive examples, negative examples and overlap guards where nearby rules might also match.

Sources and references

These references are used as provenance for review heuristics, background or related materials.

Sources/references for holistic patterns

• Wikipedia: Signs of AI writing - wikipedia_signs_ai_writing: Secondary source and community field guide for prose-level AI writing signs. It frames signs as indicators to combine with context, not deterministic proof.
• OpenAI Harmony response format - openai_harmony_docs: OpenAI documentation for Harmony message formatting, channels, tool calls and browser citation markers.
• OpenAI citation formatting guide - openai_citation_formatting_docs: OpenAI documentation for private-use citation markers, source reference IDs and optional line or block locators.
• OpenAI File Search guide - openai_file_search_docs: OpenAI documentation showing file-search citation annotations, file IDs, filenames and raw API metadata.
• Stack Overflow: OpenAI Assistants placeholder report - stackoverflow_openai_assistants_placeholder: Community report showing streamed OpenAI Assistants/File Search placeholder leaks.
• Observed ChatGPT contentReference leak - observed_chatgpt_content_reference: Empirical UI/copy-paste artefact observed in copied ChatGPT or markdown output. This is an observed/local source, not an external URL.
• Anthropic Claude citations documentation - anthropic_claude_citations_docs: Anthropic documentation for Claude citation blocks and fields.
• Google Gemini grounding documentation - google_gemini_grounding_docs: Google Gemini documentation for groundingMetadata, webSearchQueries and groundingChunks.
• Perplexity Sonar features documentation - perplexity_sonar_features_docs: Perplexity documentation showing search/citation metadata fields.
• OpenAI: Where the goblins came from - openai_goblin_article: OpenAI investigation into GPT-5.1+, GPT-5.4 Nerdy personality and GPT-5.5/Codex creature-metaphor tics.
• TidBITS Talk: ChatGPT's Goblin Obsession Evades OpenAI's Fixes - tidbits_gpt55_prompt_discussion: Discussion of Codex/GPT-5.5 prompt constraints and creature-metaphor examples.

Background/Useful

• Promptfoo: Hidden Unicode attacks - useful background on hidden Unicode and invisible prompt-injection characters. This is background for the README rather than a current HOLISTIC_SOURCE_CATALOG entry.

Changelog

See CHANGELOG.md for release notes. Record user-visible detection changes, corpus updates, scoring changes and README/test structure changes there.

License

This project is distributed under the MIT License. See LICENSE for details.