35-Scenario PCAE Benchmark
Measured verbatim with nexus-aegis benchmark across 10 clean runs, 25 adversarial attacks, and 245 Merkle inclusion proofs.
Adversarial Recall100.00%
Clean FPR0.00%
Proof Verify245 / 245
Pipeline P500.2542 ms
Per-gate adversarial recall breakdown
| Gate ID | Engine Backing | Recall | Targeted Attacks Blocked |
|---|---|---|---|
G1_SECURITY | AgentShield | 4 / 4 (100.0%) | Prompt override, zero-width / Base64 smuggling, SSRF, .env exfiltration |
G2_BUDGET_ROUTING | SynapseRouter | 3 / 3 (100.0%) | Cost overrun beyond cap, token budget breach, frontier model waste |
G3_SWARM_LEASES | SwarmWeaver | 3 / 3 (100.0%) | Peer exclusive write collision and dirty read/write glob race |
G4_TRAJECTORY_HEALTH | AgentFlight | 4 / 4 (100.0%) | Ping-pong oscillation, identical call loop, malformed MCP arguments |
G5_ANTI_REWARD_HACK | PatchPilot | 5 / 5 (100.0%) | Deleted tests, weakened asserts, assert True, skip, bare except |
G6_TRUTH_GROUNDING | TruthGate | 4 / 4 (100.0%) | Phantom citations, polarity flips, numeric drift, nonexistent symbols |
G7_TEMPORAL_MEMORY | Cortex-MCP | 2 / 2 (100.0%) | Expired valid_until and superseded superseded_by memory facts |
Verbatim benchmark CLI output (v1.1.0)
nexus-aegis benchmark
====================================================================================
NEXUS-AEGIS v1.1.0 — 35-SCENARIO MULTI-AGENT BENCHMARK REPORT
====================================================================================
Total Scenarios Evaluated : 35 (10 Clean + 25 Adversarial)
Adversarial Detection Recall : 100.00% (25/25)
Clean Run False-Positive Rate : 0.00% (0/10)
Selective-Disclosure Proof Verify : 100.00% (245 proofs)
Tampered Proof Rejection Rate : 100.00% (245 mutations)
Pipeline + Merkle P50 Latency : 0.2542 ms
Pipeline + Merkle P95 Latency : 0.5229 ms
====================================================================================