03 The 40 Checks Explained

The 40-Point QA System

Every piece of Ghost Writer content is validated against 40 checks organized into 10 blocks (A–J). Hard checks must pass; soft checks inform quality. Content passes when there are 0 hard fails and ≤3 soft fails.

---

Block A: Statistical (1–7)

Counters: Perplexity-based detectors, burstiness analysis, vocabulary uniformity.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
1	Sentence Length Variance	Hard	stdev ≥ 5	Low burstiness	Injects varied sentence lengths; prompt demands mix of fragments and long sentences
2	Vocabulary Richness (TTR)	Soft	TTR ≥ 0.45	Lexical uniformity	Type-token ratio enforced; voice profile adds domain terms
3	Hapax Legomena Ratio	Soft	≥ 0.25	Repetitive vocabulary	Words used once; human-like vocabulary diversity
4	Average Sentence Length	Soft	8–25 words	Uniform structure	Prompt specifies avg sentence length per voice
5	Short Sentence Presence	Hard	≥ 1 sentence ≤ 5 words	No fragments	Prompt: "Use fragments when they hit harder"
6	Long Sentence Presence	Soft	≥ 1 sentence ≥ 25 words	Overly simple structure	Prompt: "Mix 3-word fragments with 25–35 word complex sentences"
7	N-gram Diversity	Soft	Varied distribution	Predictable token patterns	Temperature 0.85–0.95; varied generation

---

Block B: Classifier Resistance (8–12)

Counters: Binary AI/human classifiers, model attribution.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
8	Conjunction Starters	Hard	≥ 1 paragraph starts with And/But/So	AI-typical sentence structure	Prompt: "Start at least 2 paragraphs with conjunctions"
9	Fragment Usage	Soft	Contains fragments	Uniform sentence types	Regex checks for short sentences; prompt encourages fragments
10	Parenthetical Asides	Soft	Contains () or —	No human thought injection	Prompt: "Include at least one parenthetical aside"
11	Temperature Variance	Soft	0.85–0.95	Low-temperature uniformity	Generation uses 0.85+; variants use 0.85, 0.89, 0.93
12	Model Attribution Defense	Soft	Varied patterns	Model fingerprinting	Voice-specific vocabulary, structural habits break attribution

---

Block C: Linguistic (13–18)

Counters: Phrase-based detectors, readability uniformity, lexical patterns.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
13	Phrase Blacklist	Hard	0 hits	120+ AI-detectable phrases	Blacklist in prompt; post-generation replacement if hit
14	Lexical Diversity	Soft	TTR ≥ 0.50	Vocabulary repetition	Higher TTR target; voice vocabulary
15	Readability Variance	Soft	Flesch-Kincaid 20–100	Uniform readability	Voice-driven; no fixed grade level
16	Syntactic Variety	Soft	stdev ≥ 4	Uniform syntax	Sentence length stdev check
17	Emotional Authenticity	Soft	Voice-driven	Flat tone	Tone anchors in voice profile
18	Metaphor/Analogy Presence	Soft	≥ 1	No creative comparison	Prompt: "Use at least one unexpected metaphor or analogy"

---

Block D: Watermark (19–20)

Counters: Invisible Unicode watermarks, metadata embedding.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
19	Unicode Normalization	Hard	No invisible chars	Zero-width, BOM, etc.	normalizeText() strips U+200B–U+200F, U+2028–U+202F, U+FEFF
20	Metadata Clean	Hard	None	Embedded metadata	Plain text output only; no hidden markers

---

Block E: Scoring (21–25)

Counters: Detector confidence thresholds, sentence-level analysis, plagiarism, anti-humanizer.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
21	Confidence Score Target	Soft	< 30% AI on all detectors	High AI probability	Triple-detector pipeline; regenerate/revise if fail
22	Sentence-Level Clean	Soft	No sentence > 80% AI	Per-sentence flagging	Revise failed sentences; splice back
23	Plagiarism Clear	Hard	< 5% match	Originality.ai	Generate from scratch; no copy-paste
24	Anti-Humanizer	Hard	Generated human, not paraphrased	Paraphrase detection	Content built with human patterns from scratch
25	Language Authenticity	Soft	Matches voice profile	Generic AI tone	Voice profile calibration in prompt

---

Block F: Bias (26–28)

Counters: Non-native bias exploitation, domain mismatch, length mismatch.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
26	Non-Native Bias Clear	Soft	No exploitation	Detector bias against non-native writers	Ethical constraint; no deliberate bias gaming
27	Domain Pattern Match	Soft	Uses domain vocabulary	Generic wording	Voice profile domain terms; check for presence
28	Length Optimization	Soft	Matches platform best length	Wrong length for platform	Platform spec bestLength; word count check

---

Block G: Adversarial (29–31)

Counters: Pattern diversity attacks, translation artifacts, authorship inconsistency.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
29	Pattern Diversity	Soft	stdev ≥ 6	Structural uniformity	Higher stdev target for adversarial robustness
30	Translation Proof	Soft	English-native	Translation artifacts	Native generation; no translation step
31	Authorship Consistency	Soft	Single voice throughout	Mixed styles	Voice profile maintains consistency

---

Block H: Infrastructure (32–34)

Counters: Single-detector gaming, encoding issues, platform violations.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
32	Multi-Detector Validation	Hard	GPTZero + Pangram + Originality	Single-detector optimization	All three must pass when configured
33	Plain Text Normalization	Hard	Normalized output	Hidden chars, encoding	Same normalizeText() as Block D
34	Platform Format Compliance	Hard	≤ platform max chars	Over-length content	Truncate to maxChars; email split by spec

---

Block I: Evaluation (35–37)

Counters: Self-evaluation bias, FPR exploitation, AI-only classification.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
35	Third-Party Benchmark	Soft	Validated externally	Internal-only scoring	Real detector APIs; no mock scores
36	FPR Exploitation Clear	Soft	No FPR gaming	False positive rate exploitation	Ethical constraint; no detector gaming
37	AI-Assisted Classification	Soft	AI-Assisted or Human	Pangram "AI" classification	Target: Human or AI-Assisted, not AI

---

Block J: Governance (38–40)

Counters: Disclosure gaps, audit gaps, provenance ambiguity.

ID	Name	Hard/Soft	Target	Detection Vector	How Ghost Writer Defeats It
38	Disclosure Compliance	Soft	Transparent use	Hidden AI use	Tool provides detection scores for user decision
39	Audit Trail	Soft	Logged	No traceability	QA report serves as audit trail
40	Provenance Proof	Soft	Built from scratch	Paraphrased/spun content	Content constructed with human patterns from scratch

---

Pass Logic

passed = (hardFails === 0) && (softFails <= 3)

• Hard fail: Content is revised (blacklist replacement + GPT revision)
• Soft fail: Counted; if >3, triggers revision
• Max revisions: 3 for QA; 3 for detector failures