Skip to content

COMPARATIVE_EVALUATION.md

Latest commit

 

History

History
98 lines (74 loc) · 3.65 KB

COMPARATIVE_EVALUATION.md

File metadata and controls

98 lines (74 loc) · 3.65 KB
Raw

Comparative Evaluation

Phase: 6 (Research Legitimacy) Status: Initial draft — invite criticism.

Systems Under Comparison

System Domain Model
authgate-kernel Agentic AI authority control Capability ownership graph, delegation lattice
OPA (Open Policy Agent) General-purpose policy engine Rego language, ABAC/RBAC
SELinux OS mandatory access control Type enforcement, Bell-LaPadula variant
Object-capability systems (Cap'n Proto, E, Capsicum) Language/OS capability systems Unforgeable object references
Sandbox runtimes (gVisor, Firecracker) Process/VM isolation Syscall interception, hardware virtualization

Comparison Axes

1. Formal verification

System Formally verified?
authgate-kernel Partial (Kani 19 harnesses, Lean 4 4 theorems)
OPA No — Rego interpreter is not formally verified
SELinux No — policy language is not formally verified
Object-cap Varies — E language has theoretical proofs; Capsicum is not proved
Sandbox runtimes No — host kernel interface is not formally verified

2. Agentic system suitability

System Designed for autonomous agents?
authgate-kernel Yes — primary use case
OPA No — designed for human-operated services
SELinux No — designed for OS-level isolation
Object-cap Partial — capability model maps to agents, but no delegation lattice
Sandbox runtimes No — isolates processes, not agent authority

3. Attenuation enforcement

System Enforces attenuation (child ⊆ parent)?
authgate-kernel Yes — structurally enforced at delegation
OPA No — policy language can grant any authority
SELinux Partial — via type transition rules
Object-cap Yes — unforgeable references cannot be amplified
Sandbox runtimes No — isolation only; no delegation model

4. Cryptographic attestation

System Signed verification results?
authgate-kernel Yes — ed25519 per decision
OPA No
SELinux No
Object-cap No (references are attestation by construction)
Sandbox runtimes No

5. Human-in-the-loop sovereignty

System Enforces human ownership / corrigibility?
authgate-kernel Yes — structural sovereignty flags
OPA Policy-dependent (can be configured)
SELinux No — no human principal concept
Object-cap No — no human principal concept
Sandbox runtimes No

authgate-kernel's Unique Position

authgate-kernel is the only system in this comparison that simultaneously:

  1. Enforces typed capability attenuation in a delegation lattice
  2. Provides cryptographic attestation of individual decisions
  3. Enforces human principal sovereignty as a structural invariant
  4. Targets autonomous agent systems as the primary use case
  5. Is partially formally verified

Gap vs object-capability systems: Object-cap systems (Cap'n Proto, Capsicum) have stronger theoretical foundations (confining power is unforgeable), but do not model human ownership hierarchies or provide per-decision attestation.

Gap vs OPA: OPA is more expressive (Rego supports arbitrary policy logic), but this expressiveness makes it harder to formally verify and easier to misconfigure.

Invitation for Criticism

This evaluation is preliminary. We invite the community to:

  • Identify systems not listed here that are closer competitors
  • Challenge the comparison axes (are we selecting favorable ones?)
  • Provide benchmark comparisons on enforce latency and policy expressiveness
  • Identify use cases where existing systems are strictly better