Contradiction Detection
How Exogram catches semantic contradictions before they become production incidents.
Core Thesis
AI agents do not contradict themselves deliberately. They contradict themselves because they have no mechanism to detect contradiction. The context window is too small to hold all prior statements. The model has no formal logic system. Contradiction detection provides the architectural mechanism that models lack — deterministic verification that a proposed action does not conflict with established operational state.
The Contradiction Taxonomy
Contradictions in agentic systems fall into four categories: (1) Factual contradictions — the agent asserts something that contradicts established facts in the environment. (2) State contradictions — the agent assumes a system state that does not match verified state. (3) Policy contradictions — the agent proposes an action that violates active governance policies. (4) Temporal contradictions — the agent proposes a sequence of actions that is logically impossible given the current state timeline. Each category requires different detection mechanisms, but all share the same architectural principle: compare the proposal against the environment, not against the model's internal beliefs.
Detection Through Environmental Verification
Contradiction detection in Exogram is not a separate system — it is a property of the environmental architecture. When an agent proposes an action, the control plane naturally evaluates it against the verified state graph. If the agent claims "User X has admin privileges" but the identity graph shows User X has read-only access, the contradiction surfaces through normal policy evaluation. The detection is a byproduct of governance, not a bolted-on feature.
Cross-Session Contradiction Memory
Context windows reset between sessions. Contradictions don't. An agent that deleted a resource in Session 1 may attempt to modify it in Session 2, because Session 2 has no memory of Session 1. Event-sourced memory solves this: the deletion event is permanently recorded. When the modification is proposed, the ledger proves the resource no longer exists. Cross-session contradiction detection is automatic in event-sourced architecture.
Multi-Agent Contradiction Resolution
When two agents propose contradictory actions — one wants to approve a transaction, another wants to flag it for fraud — resolution cannot be delegated to either agent. Both believe they are correct. The environment resolves the contradiction through structural precedence: fraud detection policy outranks approval workflow. The resolution is deterministic, auditable, and independent of which agent proposed first.
Frequently Asked Questions
How does Exogram detect contradictions in AI agent actions?+
Through environmental verification. Every proposed action is evaluated against the verified state graph and event-sourced history. If the proposal conflicts with established state, active policies, or recorded history, the contradiction is detected before execution.
Can contradiction detection catch hallucinations?+
Yes. Schema hallucinations (invented parameters), state hallucinations (assumed state that doesn't exist), and factual hallucinations (assertions contradicting established facts) are all caught by comparing the agent's proposal against the verified environment.
Related Architecture
Related Glossary
Related Learn Articles
Competitor Comparisons
Deploy This Architecture
Stop building AI systems without coherent operational environments. Start governing agent actions with deterministic infrastructure.