Primus Guard

Primus Guard - Pre-Execution Enforcement for Autonomous Agents Most agent systems focus on what an agent can do. Primus Guard enforces what an agent is allowed to do before execution. We extracted a deterministic policy layer from a live Primus demo system and applied it to agent actions. What it does: - Intercepts proposed agent actions - Applies deterministic policy constraints - Blocks unsafe or invalid actions before execution - Produces structured audit artifacts - Generates replayable decision records Why this matters: Current agent systems assume correct behavior. Primus removes that assumption. Every decision is: - evaluated - enforced - provable ## Live Enforcement Example Below is a real enforcement outcome from Primus Guard: ```json { "execution_blocked": true, "violation_count": 1, "violations": [ { "rule": "slippage_guard", "reason": "slippage_too_high_75_vs_50" } ], "audit_log": "C:\\PrimusQuant_Laptop\\logs\\audit.jsonl", "decision_hash": "e8890931fa4339e4696faf6ffde4a31f323e67e0236850e30f79d3c8bf101699" } ``` This decision was: - evaluated before execution - blocked deterministically - recorded in an append-only audit log - cryptographically verifiable via replay Implementation reference: - Local repo reference: primus_guard_demo - Core system reference: C:\PrimusQuant_Laptop - Audit output: logs/audit.jsonl - Replay verification: deterministic re-evaluation of the same action and config returns the same block or allow result Tested against adversarial-style conditions inspired by real trading system filters.