Trust Hierarchy & State Isolation

To ensure that industrial evaluations are both secure and non-repudiable, AgentV implements a Tiered Trust Model and Immutable State Isolation. This prevents external plugins or untrusted metrics from corrupting the evaluation trajectory.

🛡️ 3-Tier Trust Hierarchy

The harness classifies all components (Metrics, Plugins, Analyzers) into three trust tiers. This determines their level of access to sensitive session metadata and hardware telemetry.

Tier	Source	Permissions
CORE	Internal eval_runner	Full access to session_metadata, resource_telemetry, and internal buffers.
TRUSTED	Signed Member Plugins	Access to restricted system parameters. Requires cryptographic provenance.
EXTERNAL	Third-party / Ad-hoc	Restricted to sanitized agent_summary and isolated (copied) history.

Trust Gating in session.py

When a metric or plugin is invoked, the SessionManager checks its provenance:

# [AES v1.6.0] Trust Gate for System Parameters
if is_trusted:
    context_map["session_metadata"] = self.session_metadata
    context_map["resource_telemetry"] = self.resource_telemetry

---

🧬 Conditional Deep-Copy Isolation

To prevent “Trajectory Pollution,” AgentV enforces strict memory isolation. Untrusted components never receive a direct reference to the live session state.

How it works:

1. Mutable Guard: If a component is not marked as CORE and requests a mutable object (like history or actual_state), the harness automatically performs a copy.deepcopy().
2. Performance Optimization: Deep-copying is skipped for CORE components to minimize latency in high-frequency diagnostic loops.

Implementation Pattern:

def get_isolated(key, data, _params=params, _is_core=is_core):
    if key in _params and not _is_core and isinstance(data, (dict, list)):
        return copy.deepcopy(data)
    return data

---

🔒 Forensic Non-Repudiation

By combining the Trust Hierarchy with State Isolation, AgentV ensures that the final run.jsonl trace is an accurate reflection of the agent’s decisions, unpolluted by side-effects from the evaluation infrastructure itself. This satisfies the requirements for NIST AI-100-1 and GDPR Audit Trails.