POLICY CONTROL PLANE

LLMs suggest.
The Kernel decides.
Aegis enforces the boundary.

AegisAI — Policy Control Plane for AI Systems

Deterministic policy control before any tool call, state change, or external action.

Deterministic adjudication before action
Evidence-grade decision traces
Designed to support regulated deployments
See EvidenceRead ArchitectureTry the demo — adapter mode exercises real Kernel

WHY AEGISAI EXISTS

LLMs are probabilistic

They cannot be audited. Same input may yield different behavior in different contexts.

Prompts are not policies

Text instructions are suggestions, not enforceable controls. Policies are.

Governance requires evidence

Regulators and operators need determinism — and a verifiable artifact trail.

Æ

"In ancient Greece, Aegis meant an impenetrable shield — a symbol of enforced protection, not passive trust."

The name reflects the architecture: AI systems are not trusted by default. They are contained, governed, and constrained by a deterministic kernel.

AEGISAI EXECUTION BOUNDARY

LLM ROLE
Proposal generation only — unprivileged
KERNEL ROLE
Adjudication, policy, decision
GOVERNANCE
Panic mode & Break-Glass (signed)
VIOLATIONS
Hard-blocked by design, never soft-degraded

LLMs never execute actions.
All actions dispatched exclusively through Kernel adjudication.

DECISION FLOW

Upstream model / LLM
stochastic — same input may produce different output
↓ proposes raw text output
LLM Adapter
normalization → typed ProposalSpec (schema-validated)
↓ typed ProposalSpec
AegisAI Kernel / PolicyEngine
deterministic · auditable · traceable
↓ decision outcome
ALLOWWARNREQUIRE_CONFIRMBLOCK
↓ always — regardless of outcome
DecisionTrace
session_id · custody_hash (SHA-256) · audit_log
↓ if AEGIS_SIGNING_KEY configured
DecisionTraceAttestation
Class A · Ed25519 signed · externally verifiable

EVIDENCE-GRADE DECISIONING

Every adjudication. Verifiable.

No decisions happen off the record.

SHA-256
Custody hash
Every DecisionTrace
Ed25519
Attestation
Class A — externally verifiable
Chain
Tamper-evident
Integrity by design
CLI
Reproducible
Deterministic + CLI verification

"Logs can be deleted. Evidence cannot."

WHAT AEGISAI PROVIDES

Kernel Sovereignty

Deterministic control loop. LLM output normalized before reaching the kernel. The kernel adjudicates; the model does not.

Policy Engine

DSL-governed rules: ALLOW, WARN, REQUIRE_CONFIRM, BLOCK. No eval(). Conditions validated at load time.

Cryptographic Records

SHA-256 custody hash on every trace. Ed25519-signed attestation for Class A governance artifacts when configured.

Benchmark Evidence

7 canonical governance scenarios. Reproducible artifact. CI gate: exits 1 on any security failure.

Designed to support record-keeping, human oversight, and auditability in regulated AI deployments.

See EvidenceRequest a Pilot