Symbolic Governance for Probabilistic Intelligence

Why Governable AI Requires Symbolic Structure Without Reverting to Symbolic AI

FERZ Canonical White Paper Series | December 2025

The AI Governance Crisis

The rapid deployment of probabilistic AI systems—large language models, autonomous agents, tool-calling workflows—has created a governance crisis that existing mechanisms cannot address. These systems produce outputs through statistical inference over learned parameters. They cannot reliably explain why a specific output was generated. They cannot guarantee that the same input will produce the same output. They cannot verify that their outputs comply with external constraints.

The standard response has been to pursue alignment: training models to behave safely, adding guardrails through prompts, inserting humans into approval workflows. These approaches treat governance as a property of the AI system itself. They fail because probabilistic systems cannot provide the deterministic guarantees that governance requires. A model that is 99.9% likely to comply is not a governed model—it is an ungoverned model that usually behaves well.

The Core Insight: Intelligence vs. Governance

This paper argues that governable AI requires a fundamental separation between intelligence and governance. Neural systems excel at generating outputs: hypotheses, recommendations, proposed actions. But the authorization of those outputs—the decision to allow an action to take effect—must occur through a different mechanism entirely.

That mechanism must be deterministic, replayable, and verifiable. It must carry proof, not explanation. It must fail closed when verification cannot be completed.

"Intelligence proposes; governance verifies. The system that generates candidates is not the system that authorizes them."

The FERZ Architecture

FERZ introduces a deterministic envelope: an architectural boundary that separates probabilistic generation from deterministic enforcement. Inside the envelope, neural systems operate as they normally do—generating outputs, proposing actions, producing recommendations. Outside the envelope, nothing happens until a deterministic verification process confirms that the proposed action is authorized.

Proof-Carrying Decisions

Central to FERZ is the concept of Proof-Carrying Decisions (PCDs): decision artifacts that contain not just the decision outcome, but the evidence required to verify that outcome independently. A PCD includes the action schema, authority claims, applicable policies, required evidence, constraint evaluations, and cryptographic signatures. No proof, no action.

Fail-Closed Enforcement

FERZ implements fail-closed enforcement: if verification fails, execution halts. When verification cannot establish authorization, the default is denial. Actions do not proceed when evidence is missing, when constraints cannot be evaluated, or when authority cannot be established. Uncertainty does not grant permission.

Replayable Verification

Every governed decision must be replayable. Given the same inputs and the same context, the same verification produces the same result. This is essential for audits, for dispute resolution, for incident investigation. If a decision cannot be replayed, it cannot be governed.

Why This Is Not "Symbolic AI 2.0"

A superficial reading might conclude that FERZ represents a return to symbolic AI—rules instead of neural networks, logic instead of statistics. This interpretation misses the point entirely.

Symbolic AI failed as a theory of intelligence—it could not handle open-world reasoning, could not scale knowledge acquisition, could not match neural systems on perception and language. But symbolic AI succeeded as infrastructure. The properties that made it inadequate for cognition—determinism, inspectability, composability—make it essential for governance.

FERZ uses symbolic structure for what it is good at: verification. We leave intelligence to the systems that are good at it.

Practical Implications

For enterprises: AI systems can be deployed in regulated environments with verifiable compliance guarantees. The compliance posture is defined by architecture, not by hope.

For regulators: Enforcement can shift from documentation review to mechanical verification. Instead of asking "what are your policies," regulators can ask "show me the PCDs for decisions in scope."

For auditors: Instead of reconstructing decisions from logs and interviews, auditors replay the exact verification that occurred at decision time. The audit produces definitive findings, not probabilistic assessments.

For the AI industry: A path to deployment in high-stakes domains that does not require waiting for alignment to be solved.

The Six Canonical Principles

The full white paper establishes six canonical governance principles that form the foundation of FERZ:

Intelligence proposes. Neural systems generate candidates for action—proposals are inputs to governance, not decisions.
Governance verifies. Authorization must be performed by systems that are deterministic, inspectable, and verifiable.
Decisions must carry proof. A decision without proof is not a governed decision.
Verification must be replayable. A decision that cannot be reconstructed cannot be audited.
Enforcement must fail closed. When verification cannot establish authorization, the default is denial.
Authority must be explicit. Every decision must trace to a documented source of authority.

Download the Full White Paper

The complete 39-page white paper includes formal definitions, worked examples in healthcare and financial services, threat models, falsifiability criteria, and mappings to FERZ technical standards.

Document ID: FERZ-WP-2025-001
Version: 1.2.0 (Canonical)
Download: https://doi.org/10.5281/zenodo.18072966