1. AI as the Advisory Layer LLMs are useful for understanding intent, grouping requirements, explaining alternatives, and suggesting the next question. A proposal, however, should not automatically become a business decision.
The system validates proposed products, modules, and integrations independently.
2. Canonical Project State Confirmed answers should live in structured state independent from conversation text. This enables reproducible evaluation, version comparison, session recovery, and a clear record of decisions confirmed by the user.
Conversation becomes an interface to state rather than the state itself.
3. Deterministic Evaluation Scope, commercial rules, dependencies, and constraints should pass through deterministic backend logic rather than being decided by AI on its own. The same input under the same rule version should produce the same result.
AI can explain the outcome but should not silently rewrite the rules.