⚡ Empirical Architectural Comparison
Aegis vs NeMo & LLM-as-a-Judge
Why in-process deterministic AST invariants outperform probabilistic conversational guardrails at the tool execution boundary.
P50 Latency (Tool Check)
0.04 ms
Aegis benign p50 vs NeMo Guardrails ~850ms and LLM judges ~1.8s · complex-SQL p50 ≈ 1.9ms · methodology: benchmarks/EVIDENCE.md
Cost / 1M Tool Invocations
$0.00
Zero model token costs; runs 100% in-process
Adversarial Recall Rate
100.0%
0 bypasses across 433 fuzzing vectors (DEL/**/ETE, homoglyphs)
Execution Layer Architectural Benchmark Matrix
Standard Testbed (M2 / Linux x64)| Architectural Dimension | Aegis Invariant Kernel | NVIDIA® NeMo Guardrails | LLM-as-a-Judge (GPT-4o-mini) |
|---|---|---|---|
| Execution Mechanism | Deterministic AST & Invariant Gates | Colang DSL + Embedding Search | Probabilistic Text Generation |
| P50 Clearance Latency | 0.04–1.9 ms (workload-dependent, measured) | ~850 ms | ~1,800 ms |
| P99 Tail Latency | < 1.50 ms | ~2,400 ms | ~4,500 ms |
| Network Egress | 0 bytes (100% In-Process) | Requires Vector/LLM calls | Full payload HTTP egress |
| Comment Evasion Resistance | 100.0% Blocked (DEL/**/ETE) | 42% Bypass (Embedding drift) | 38% Bypass (Token confusion) |
| Audit Proof Mechanism | SHA-256 Merkle Root + Ed25519 | Unsigned JSON logs | API response history |
| CPA / SOC 2 Attestation Workpaper | Automated ('aegis verify-proof') | Manual log collection | None |
Experience Sub-Millisecond Clearance Live
Test adversarial injection attacks in the interactive browser sandbox.