QILAHK™ / AI & TRUST

Define a Proof Surface for an AI System

Scope an AI assessment around the system, its permissions, evidence, and intended use.

October 12, 2026 · QILAHK Journal

Assess a defined system

An AI assessment needs more than a model name. The relevant system may include a model version, application logic, retrieval sources, tools, permissions, deployment configuration, and operating context. State which of these components the claim covers.

QILAHK’s Proof Surface concept provides a way to frame that boundary: what would have to be demonstrated for this specific claim to be defensible? The answer depends on the intended use and consequences of error.

Make claims testable

“The AI is trustworthy” is too broad to evaluate directly. More useful claims concern defined behavior or controls: whether a tool invocation requires authorization, whether retrieval sources are restricted as specified, or whether a release met documented evaluation criteria.

Define the evaluation conditions and retain representative results, limitations, and failure observations. A successful demonstration under one set of inputs does not establish every behavior in every operating context.

Track provenance and permissions

Identify the model and application versions, configuration, source provenance where available, and permissions granted to connected tools. Distinguish the ability to generate an answer from the authority to take an action.

Evidence should also show how changes are managed. A model replacement, new data source, expanded tool permission, or changed instruction can affect the basis of a prior assessment even when the interface looks identical.

Keep reliance within the boundary

Document what the evaluation establishes, what remains uncertain, and what requires human review. A trust determination should reflect the evidence and use conditions rather than implying universal correctness.

A practical starting record includes the intended task, system components, authorized actions, evaluation criteria, evidence references, exclusions, and revalidation triggers. This creates a basis for accountable review as the AI system changes, while making the limits of the assessment clear to those relying on it.

Continue exploring the foundations in Cybersecurity Has Security Controls. What It Lacks Is a Trust Layer.