Audit Framework

The AISYSTEMS Audit Framework is the methodological authority of AISYSTEMS s. r. o.

It represents a normative, deterministic, and observe-only layer designed for the technical observation of AI system behavior.

The Audit Framework:

  • does not modify configuration of target systems,
  • does not alter their runtime conditions,
  • does not optimize their behavior,
  • defines the rules for generating technical evidence,
  • does not issue evaluations or verdicts.

The Audit Framework defines methodology. It does not perform execution.

How It Works

The audit is based on a reproducible and deterministic principle:

  • one input → one output,
  • no context retention between runs,
  • application of explicitly defined rules,
  • strict separation of measurement from interpretation.

Each run is bound to a specific version of the methodology and explicitly anchored within an AuditPlan.

The output is a structured technical record (Audit Evidence), not an evaluation or decision.

AISYSTEMS SENTINEL

AISYSTEMS SENTINEL is the deterministic execution implementation of the Audit Framework.

It operates as a neutral measurement device that:

  • executes an approved AuditPlan,
  • interacts with the target system as a black box,
  • applies rules defined by the Framework,
  • generates cryptographically verifiable technical artifacts.

SENTINEL has no authority to modify methodology or Framework rules. It is an execution layer, not an authority.

Technical documentation: auditumelejinteligencie.sk/#sentinel

Validation

The execution layer has been technically validated in:

  • a local open-weight environment,
  • an external API environment.

Testing included repeated runs of identical scenarios under identical conditions to verify reproducibility.

Observed phenomena included:

  • cases of ambiguous or implicit AI identity,
  • language-dependent behavioral differences,
  • differences between local and cloud deployments.

Validation confirmed:

  • determinism of audit logic,
  • a fully stateless execution model,
  • architectural independence of the methodology from deployment type.

Language Context

Audit scenarios were executed in both Slovak and English.

Observations indicate that in non-dominant language contexts there may be a higher occurrence of:

  • inaccurate claims,
  • identity drift,
  • hallucinatory responses.

These phenomena are recorded as technical artifacts without evaluative conclusions.

Boundaries

The AISYSTEMS Audit Framework produces exclusively technical evidence based on an observe-only methodology.

It does not hold regulatory, certification, or legal authority. It does not issue compliance assessments or binding decisions.

Its role is measurement. Decision-making belongs to people and institutions.