Architecture

Architecture

flowchart LR
    A[Versioned Agent Manifest] --> B[Change Fingerprinting]
    B --> C[Assurance / Impact Graph]
    C --> D[Selective Test Planner]
    D --> E[Execution and Oracles]
    F[DeltaBench Scenarios] --> C
    F --> E
    E --> G[Runtime Evidence and Traces]
    G --> H[Certificate Status]

Loading

Versioned manifest

Captures security-relevant state such as model identity, prompts, tools, schemas, permissions, policies, memory configuration, dependencies, and runtime controls.

Change fingerprinting

Records which security-relevant elements differ between the certified and candidate configurations.

Assurance / impact graph

Links changed components to assurance claims and candidate tests.

Selective test planner

Constructs a scoped suite from impacted tests plus risk-triggered sentinels. It escalates when impact cannot be bounded conservatively.

Execution and oracles

Runs selected scenarios and evaluates task behavior, policy violations, containment, attack success, and regression outcomes.

Certificate state

Evidence remains configuration-scoped and traceable to the exact change, model profile, mutation, test, and repetition.