Add ZeroGPT

Author: MADEVAL

EVAL.md

@@ -1,5 +1,7 @@
 # EVAL - Quality Evaluation Framework for HUMAN-AI
 
+> **NEW in v3.0:** ZeroGPT external validator integration. Run `scripts/run-benchmark.ps1` with a ZeroGPT API key to get independent AI detection scores.
+
 > Load this prompt alongside SKILL.md output to score the humanization result.
 > The evaluator is a separate prompt to avoid self-assessment bias.
 
@@ -166,3 +168,70 @@ HUMAN-AI OUTPUT:
 ```
 
 The evaluator must have access to `shared/tone-profiles.md` and `shared/rhythm-tables.md` for accurate scoring.
+
+---
+
+## External Validation — ZeroGPT API
+
+For **independent** validation (not LLM self-assessment), use ZeroGPT's external AI detection API.
+
+### Prerequisites
+
+1. ZeroGPT API key (register at [zerogpt.com](https://www.zerogpt.com) → Dashboard → API)
+2. Set environment variable: `ZEROGPT_API_KEY="your-key"`
+
+### Quick single-text check
+
+```bash
+# PowerShell
+.\scripts\zerogpt-detect.ps1 -Text "Text to analyze"
+
+# Bash
+./scripts/zerogpt-detect.sh --text "Text to analyze"
+
+# From file
+.\scripts\zerogpt-detect.ps1 -File "path/to/text.md"
+```
+
+### Full benchmark run
+
+```bash
+# Dry run (see what would be tested)
+.\scripts\run-benchmark.ps1 -SkipApi
+
+# Full run with ZeroGPT API
+$env:ZEROGPT_API_KEY = "your-key"
+.\scripts\run-benchmark.ps1
+```
+
+Results are saved to `tests/benchmark/zerogpt-results.json`.
+
+### Interpreting ZeroGPT scores
+
+| ZeroGPT Score | Meaning | Action |
+|--------------|---------|--------|
+| 0-10% | Human-written | Pre-flight guard should have STOPPED |
+| 10-25% | Mostly human | Acceptable output |
+| 25-50% | Mixed signals | Some AI patterns remain |
+| 50-80% | Likely AI | Re-run pipeline with aggressive cleanup |
+| 80-100% | Heavy AI | Pipeline did not work — adjust parameters |
+
+### Combined Evaluation Flow
+
+```
+1. Run skill → get self-evaluated [QUALITY: XX/100]
+2. Run ZeroGPT on output → get external AI probability
+3. Run EVAL.md LLM evaluator → get independent 5-metric score
+4. Triangulate: if all three agree → high confidence in quality
+```
+
+### Exit codes
+
+| Code | Meaning |
+|------|---------|
+| 0 | Text classified as human / mixed |
+| 10 | Text classified as AI-generated |
+| 1 | Input error |
+| 2 | Missing API key |
+| 3 | API connection error |
+| 4 | API returned error |

README.md

@@ -91,6 +91,18 @@ natural-skill/
 │   ├── specificity-ladder.md       ← Abstraction → concrete framework
 │   ├── rhythm-tables.md            ← Sentence flow parameters
 │   └── language-template.md        ← Template for adding new languages
+├── scripts/                      ← Automation scripts
+│   ├── validate.ps1                ← Integrity checker (PowerShell)
+│   ├── validate.sh                 ← Integrity checker (Bash)
+│   ├── zerogpt-detect.ps1          ← ZeroGPT AI detection (PowerShell)
+│   ├── zerogpt-detect.sh           ← ZeroGPT AI detection (Bash)
+│   └── run-benchmark.ps1           ← External benchmark runner
+├── tests/benchmark/               ← Evaluation dataset
+│   ├── annotations.json
+│   ├── zerogpt-results.json        ← ZeroGPT external validation results
+│   ├── ai-texts/  (9 languages)
+│   ├── human-texts/
+│   └── edge-cases/
 ├── scenarios/                      ← Task-specific playbooks
 │   ├── full-rewrite.md             ← Default: all 6 stages
 │   ├── blog-post.md                ← Blog post humanization
@@ -189,7 +201,35 @@ Here's what the four good ones have in common. And why the other eight failed.
 - Any capable LLM with a system prompt / custom instructions field
 - For full skill functionality: a skill system that loads files from a folder (OpenCode, Claude Code)
 - For standalone: copy `SKILL.md` as system prompt - contains complete pipeline
-- No API keys, no tools, no dependencies - pure prompt engineering
+- **External validation:** ZeroGPT API key (optional) — for independent AI detection scoring via `scripts/run-benchmark.ps1`
+
+## Quality Validation
+
+HUMAN-AI provides **three layers** of quality verification:
+
+| Layer | Method | Independence | How to run |
+|-------|--------|-------------|------------|
+| 1. Self-eval | Stage 5.6 Quality Score | LLM self-assessment | Built into pipeline |
+| 2. EVAL.md | 5-metric LLM evaluation | Separate LLM call | `EVAL.md` prompt |
+| 3. ZeroGPT | External API detection | Fully independent | `scripts/run-benchmark.ps1` |
+
+**ZeroGPT external validation:**
+
+```bash
+# Set API key
+$env:ZEROGPT_API_KEY = "your-key"
+
+# Single text check
+.\scripts\zerogpt-detect.ps1 -File "output.md"
+
+# Full benchmark run (all test texts)
+.\scripts\run-benchmark.ps1 -MaxTexts 10 -DelaySeconds 2
+
+# Dry run (see what would be tested)
+.\scripts\run-benchmark.ps1 -SkipApi
+```
+
+Results saved to `tests/benchmark/zerogpt-results.json`.
 
 ---
 

SKILL.md

@@ -367,6 +367,8 @@ If QUALITY SCORE < 75: identify the 2 lowest-scoring checks. Re-run ONLY the rel
 
 After the single re-loop, output the final score regardless. Stop after 2 passes total.
 
+**External validation:** For independent verification, run ZeroGPT detection via `scripts/zerogpt-detect.ps1` or `scripts/run-benchmark.ps1`. See `EVAL.md` — External Validation section for details. The self-evaluation score is an LLM self-assessment. ZeroGPT provides an independent external check.
+
 ---
 
 ## WHEN NOT TO APPLY
@@ -495,11 +497,15 @@ natural-skill/
 ├── .gitignore
 ├── scripts/
 │   ├── validate.ps1                ← Integrity checker (PowerShell)
-│   └── validate.sh                 ← Integrity checker (Bash)
+│   ├── validate.sh                 ← Integrity checker (Bash)
+│   ├── zerogpt-detect.ps1          ← ZeroGPT AI detection (PowerShell)
+│   ├── zerogpt-detect.sh           ← ZeroGPT AI detection (Bash)
+│   └── run-benchmark.ps1           ← External benchmark runner (ZeroGPT)
 ├── .github/workflows/
 │   └── validate.yml                ← CI pipeline
 ├── tests/benchmark/                ← Evaluation dataset
 │   ├── annotations.json
+│   ├── zerogpt-results.json        ← ZeroGPT external validation results
 │   ├── ai-texts/  (9 languages)
 │   ├── human-texts/
 │   └── edge-cases/