Safety Architecture
Verification vs Prompt Engineering
Prompt engineering is probabilistic. Constitutional AI verification is deterministic. For enterprise compliance, the difference is not philosophical โ€” it is legal.
The Critical Difference
Prompt engineering adds tokens that make certain outputs more probable. A model can still ignore them โ€” 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 โ€” always. No model can override this.
Prompt Engineering
Instructions to the model
- โ€ขWorks at the model level (text input)
- โ€ขProbabilistic โ€” model may ignore
- โ€ขDifferent prompts needed per LLM
- โ€ขNo audit trail built-in
- โ€ขVulnerable to context dilution (long conversations)
- โ€ขVulnerable to prompt injection attacks
โ“ Excellent for task formatting & style
Constitutional AI Verification
Constraints on the system
- โ€ขWorks at the system level (around the model)
- โ€ขDeterministic โ€” mathematically guaranteed
- โ€ขOne constraint set, works across all 7 HexaCore models
- โ€ขFull audit trail (RCTDB + JITNA packet log)
- โ€ขPer-packet validation โ€” no context dilution
- โ€ขJITNA Normalizer strips injection attempts pre-LLM
โ“ Required for regulated industry compliance
Capability Matrix
| Capability | Prompt Engineering | Constitutional Verification |
|---|---|---|
| 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 |
Read the Full Analysis
Detailed explanation of 4 prompt engineering failure modes and FDIA's 3-level verification