Semantic Guardrails
Why it matters
Traditional guardrails use rules-based filters (blocklists, if-then logic) that can stop toxic language or PII leaks but cannot prevent strategic misalignment. Semantic guardrails use a lightweight policy agent to evaluate the semantics of an output — whether a sales agent promising an unplanned feature to close a deal, or a customer success agent offering refunds when the strategy calls for product education. Unlike security guardrails, semantic guardrails enforce strategic coherence, not just safety.
How Stratafy addresses this
Stratafy implements semantic guardrails as a meaning-based policy layer — not keyword filters or rule-based blocklists. Foundation-level constraints (values, principles, risk tolerance) are queryable by AI agents as operational boundaries, and a governance model enforces authority levels across the strategy architecture.
Meaning-based evaluation, not rule-based filtering
Traditional guardrails stop toxic language or PII leaks but cannot prevent strategic misalignment. Semantic guardrails catch a sales agent promising an unplanned feature to close a deal, or a customer success agent offering refunds when the strategy calls for product education. These are semantically misaligned but would pass any keyword filter.
Foundation as queryable constraint layer
Values, principles, and risk tolerance are not inspirational wall art — they are structured, queryable data that AI agents check before acting. When a decision conflicts with a stated value, the system flags it. This turns organisational identity from decoration into operational infrastructure.
Governance by level across the strategy architecture
Each layer has its own governance model enforced through semantic guardrails. Foundation changes require board approval on an annual cadence. Strategy changes require leadership consensus quarterly. Initiatives need budget holder approval monthly. The higher the level, the more deliberation required for change.
Trust spectrum as operational guardrails
Five modes from fully autonomous to AI informs, each defining what AI can and cannot do. Agents in autonomous mode handle metric collection and status updates. Foundation-level changes require humans in full control. Each mode is itself a guardrail — structurally preventing agents from exceeding their authority.