Safety Architecture
Verification vs Prompt Engineering
Prompt engineering is probabilistic. Constitutional AI verification is deterministic. For enterprise compliance, the difference is not philosophical — it is legal.
The Critical Difference
Prompt engineering adds tokens that make certain outputs more probable. A model can still ignore them — especially on long conversations, adversarial inputs, or after fine-tuning. Constitutional AI constraints are evaluated by the system, not the model. When A=0 in the FDIA equation, F=0 — always. No model can override this.
Prompt Engineering
Instructions to the model
- • Works at the model level (text input)
- • Probabilistic — model may ignore
- • Different prompts needed per LLM
- • No audit trail built-in
- • Vulnerable to context dilution (long conversations)
- • Vulnerable to prompt injection attacks
✓ Excellent for task formatting & style
Constitutional AI Verification
Constraints on the system
- • Works at the system level (around the model)
- • Deterministic — mathematically guaranteed
- • One constraint set, works across all 7 HexaCore models
- • Full audit trail (RCTDB + JITNA packet log)
- • Per-packet validation — no context dilution
- • JITNA Normalizer strips injection attempts pre-LLM
✓ Required for regulated industry compliance
Feature Comparison Table
| Capability | Prompt Engineering | Constitutional AI |
|---|---|---|
| Prevents prompt injection | ||
| Deterministic output blocking | ||
| Works identically across all LLMs | ||
| Built-in audit trail | ||
| Scales with context window | ||
| Enables multi-model consensus | ||
| Quick iteration for task style/format | ||
| No code changes needed | ||
| Compliance documentation | ||
| PDPA Section 33 explainability |
Yes Partial No