Why this benchmark
Everyone says AI is powerful. But what wasn't said: 'powerful at what?'
Adam is designed for 7 specific tasks. GPT-4 is designed for everything. The question: when does Adam win, when does GPT-4 win?
Methodology
- 7 tasks, each with 100 scenarios
- Adam = the full version with all 12 layers
- GPT-4 = gpt-4-1106-preview via API
- Each scenario = one human judge + rubric
- Score = accuracy 0-100
The 7 tasks
1. Egyptian dialect decoding
Task: convert a question in Egyptian colloquial to JSON structured intent
| Model | Score |
|-------|-------|
| Adam Prism | 94% |
| GPT-4 | 87% |
Why Adam wins: Adam was trained on 742 real Egyptian SCADA conversations. GPT-4 understands, but with less accuracy on technical jargon.
2. Yokogawa DCS fault codes
Task: explain fault code + first 3 troubleshooting steps
| Model | Score |
|-------|-------|
| Adam Prism | 96% |
| GPT-4 | 71% |
Why Adam wins: 891 technical documents. GPT-4 knows Yokogawa, but doesn't know the specific fault taxonomy for a given model.
3. SCADA I/O point reading
Task: read a reading from an I/O point and interpret it
| Model | Score |
|-------|-------|
| Adam Prism | 92% |
| GPT-4 | 76% |
4. Legal regulatory text (Egyptian)
Task: extract obligations from a legal text
| Model | Score |
|-------|-------|
| Adam Prism | 88% |
| GPT-4 | 93% |
Why GPT-4 wins: GPT-4 has a broader corpus on law. Adam specializes in industrial regulatory.
5. Creative writing (Arabic)
Task: write a short story from a prompt
| Model | Score |
|-------|-------|
| Adam Prism | 72% |
| GPT-4 | 91% |
Why GPT-4 wins: Adam is not designed for creative writing. Explicit about its scope.
6. Code generation (Python)
Task: write a Python function from a description
| Model | Score |
|-------|-------|
| Adam Prism | 78% |
| GPT-4 | 89% |
7. Air-gapped operation
Task: run the system without internet
| Model | Score |
|-------|-------|
| Adam Prism | 100% |
| GPT-4 | 0% (needs API call) |
Conclusion
Adam ≠ a replacement for GPT-4. Adam = specialized sovereign AI.
| Use case | Winner |
|----------|--------|
| Arabic industrial context | Adam |
| Legal/regulatory breadth | GPT-4 |
| Creative writing | GPT-4 |
| Air-gapped / sovereign | Adam |
| Code generation | GPT-4 |
Adam has no problem saying 'not my job'. If you ask it to write a story, it says: 'this is not my specialization.' Explicit, not a claim.
Full benchmark
Available on GitHub: /benchmarks/sovereign-ai-2026/
