02 Architecture Deep Dive

Architecture Deep Dive

Ghost Writer uses a 5-stage pipeline to generate content that passes AI detection. This document explains each stage, data flow, the revision loop, and how detection APIs are orchestrated.

---

The 5-Stage Pipeline

Every piece of content flows through five sequential stages:

┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐    ┌──────────┐
│ Profile  │───▶│  Writer  │───▶│    QA    │───▶│ Adapter  │───▶│  Polish  │
└──────────┘    └──────────┘    └──────────┘    └──────────┘    └──────────┘

Stage 1: Profile

Purpose: Load voice profile, set perplexity/burstiness targets, identify platform constraints.

What happens:

• Voice profile is loaded (built-in or custom)
• Perplexity target range is set (e.g., 30–45 for john-williams)
• Burstiness target is set (high, medium, or low)
• Platform spec is loaded (character limits, format, structure)
• Phrase blacklist (120+ banned phrases) is injected into the prompt

Output: A fully parameterized generation context.

---

Stage 2: Writer

Purpose: Generate content using GPT-5.4 with the Ghost Protocol system prompt.

What happens:

• System prompt includes: Ghost Protocol laws, voice profile, platform spec, detection evasion rules
• Temperature: 0.85–0.95 (varies per variant for diversity)
• Model: GPT-5.4
• Max tokens: 4096

Ghost Protocol laws in the prompt:

1. Controlled Chaos — Deliberate imperfection: fragments, unexpected metaphors, varied paragraph length
2. Signature Voice — Write as the defined persona (tone, vocabulary, structure)
3. Invisible Architecture — Detection evasion baked into generation, not post-processing

Output: Raw generated text.

---

Stage 3: QA Engine

Purpose: Validate content against the 40-point QA system.

What happens:

• 40 checks run across 10 blocks (A–J)
• Hard checks must pass; soft checks inform quality
• Pass criteria: 0 hard fails, ≤3 soft fails
• If failed: blacklist hits are replaced, then a revision prompt is sent to GPT

Revision loop:

1. Run QA checks
2. If blacklist hits: replace with random replacement from phrase list
3. If hard fails or >3 soft fails: send revision prompt to GPT with failed check details
4. Re-run QA on revised content
5. Repeat up to 3 times

Output: QA-passing content or best-effort after max revisions.

---

Stage 4: Adapter

Purpose: Format content for the target platform.

What happens:

• Text is normalized (Unicode, smart quotes, em-dashes)
• Truncation applied if over platform max (e.g., LinkedIn 3000 chars)
• Platform-specific formatting:
  • Email: Split into subject, preheader, body (subject <60, preheader <90)
  • Others: Plain text or markdown per platform spec

Output: Platform-ready content.

---

Stage 5: Polish

Purpose: Human-pass simulation with small edits.

What happens:

• Detection APIs run in parallel (GPTZero, Pangram, Originality)
• If any detector fails: sentences with >80% AI probability are identified
• Failed sentences are rewritten via a focused revision prompt
• Content is re-checked by detectors
• Up to 3 revision cycles for detector failures

Output: Final content with detection scores.

---

Agent Roles

Role	Responsibility
Profile Loader	Loads voice + platform spec, builds prompt context
Writer (GPT-5.4)	Generates content under Ghost Protocol
QA Engine	Runs 40 checks, triggers revision loop
Adapter	Applies platform formatting rules
Polish / Detector	Calls external APIs, revises flagged sentences

---

Data Flow

Topic + Context + Voice + Type
         │
         ▼
   buildGhostPrompt()
         │
         ▼
   callGPT54(prompt, temp)
         │
         ▼
   normalizeText() → formatForPlatform()
         │
         ▼
   runQAChecks() ──▶ [fail] ──▶ revise ──▶ loop
         │
         ▼ [pass]
   callDetectors() ──▶ [fail] ──▶ reviseFailedSentences() ──▶ loop
         │
         ▼ [pass]
   Return { content, qa, detection, stats }

---

Revision Loop

QA Revision

When QA fails:

1. Blacklist replacement: Each blacklisted phrase is replaced with a random alternative from the replacement list
2. Structural revision: If hard fails or >3 soft fails remain, a revision prompt is sent:
   • Lists failed checks with targets
   • Asks GPT to fix only those issues
   • Temperature increased by 0.05 for more variation

Detector Revision

When a detector flags content:

1. GPTZero sentences with AI probability >80% are collected
2. A focused prompt rewrites only those sentences
3. Rewritten sentences are spliced back into the content
4. Content is re-normalized and re-checked by all detectors

---

Detection APIs in Parallel

All three detectors are called with Promise.all():

// Pseudocode
const [gptzero, pangram, originality] = await Promise.all([
  fetch(gptzeroUrl, { body: text }),
  fetch(pangramUrl, { body: text }),
  fetch(originalityUrl, { body: text }),
]);

Pass criteria:

• GPTZero: completely_generated_prob < 0.30
• Pangram: prediction is Human or AI-Assisted (not AI)
• Originality.ai: ai score < 0.30, plagiarism < 0.05

Content passes only when all configured detectors pass.

---

Platform Formatting

The Adapter applies rules from PLATFORM_SPECS:

Rule Type	Example
Truncation	LinkedIn: first 140 chars visible before "see more"
Hashtags	LinkedIn: max 3; Instagram: max 5
Structure	Reddit: TL;DR if >300 words
Format	Email: subject + preheader + body; Blog: H2/H3, meta <160
Character limits	Twitter: 280 (25K premium); Threads: 500

See Platform Adapters for the full spec per platform.