ADR-002: Separation of Reasoning, Learning, and Governance

Status: Accepted
Date: 2025-12-27

Context

In safety-critical engineering domains, a single AI paradigm should not combine reasoning, learning, and decision authority. Mixing these responsibilities reduces auditability and increases risk.

Decision

CLARISSA separates: - LLM reasoning & interaction (planning, explanation, translation), - Reinforcement learning (skill acquisition for action sequences), - Neuro-symbolic governance (rules, constraints, escalation, approvals).

Consequences

Positive

• Clear bounded autonomy and traceability.
• Governance can be audited independently from learning.

Negative

• More components and interfaces to maintain.

Neutral / Open

• Specific governance policies will mature via additional ADRs.