Architecture Overview

Separation of cognition (model) and control (kernel).

Control Flow Architecture

User / LLM

↓ (suggests intent)

AegisAI Kernel

+ Policy Engine

← deterministic, auditable, sovereign

↓ (decides outcome)

ALLOW

WARN

CONFIRM

BLOCK

↓ (if allowed)

Tool / Action Execution

↓ (always)

🔒 Audit Trace (SHA-256)

LLM cannot execute actions. Only the Kernel can.

🔐

The model maps natural language to typed intent — nothing more. It produces structured requests but has no capability to execute them.

👑

The Kernel is the sole authority for decisions and action gating. It runs deterministic policy evaluation on every request.

📊

Four possible outcomes: ALLOW (proceed), WARN (proceed + log), REQUIRE_CONFIRM (wait for human), BLOCK (deny).

🔗

Every decision generates a SHA-256 hash linked to the previous record. Tamper-evident, forensic replay possible.

{
  "tool_id": "send_payment",
  "parameters": {...},
  "correlation_id": "ulid",
  "timestamp_utc": "ISO8601"
}

{
  "status": "BLOCKED",
  "reason_code": "FINANCIAL_TX_LIMIT",
  "audit_hash": "sha256:...",
  "correlation_id": "ulid"
}