🖥️
Best Viewed on Desktop
The technical documentation and app previews are optimized for desktop viewing.
← Back to Home🏗️
Architecture
Four-layer architecture implementing the Semantic Refinement Pattern. Each layer has distinct responsibilities and clean interfaces.
L4
LLM Refinement
Triage → Correlator → AttackChain → Exploit → Remediation → Summary
L3
Schema Normalization
φ: Tool Output → Finding(id, type, severity, location, description, evidence)
L2
Parallel Tool Execution
Slither ∥ Mythril ∥ Echidna ∥ Medusa ∥ Hevm ∥ Aderyn ∥ Halmos ∥ Manticore ∥ Heimdall ∥ Panoramix ∥ 12 drain modules
L1
Pure Dispatcher
Ray/ProcessPool orchestration — NO domain logic. Manages lifecycles, resources, result collection.
↑ Input: Source Code / Bytecode
Pure Dispatcher (Definition 1)
✅ Properties of a Pure Dispatcher
An orchestration component is pure if it: (1) Contains no domain-specific logic, (2) Makes no decisions based on analysis content, (3) Treats all tools uniformly through a common interface, (4) Could be reused for entirely different analysis domains. Enables tool substitution, domain transfer, and independent testing.
Speedup Analysis (Proposition 1)
Speedup = Σᵢ tᵢ / (maxᵢ tᵢ + ε)
Follows from Amdahl's law when tools execute independently. ε = orchestration overhead. For similar times → speedup ≈ n. For heterogeneous times → bounded by slowest tool.
Sequential
162 min
Parallel (28)
23.1 min
Speedup
7.0×
Full Pipeline (Algorithm 2)
Zentinel-audit v4.3 Analysis PipelinePseudocode
Require: Target contract path P
Ensure: Analysis report R with verified exploits
tools_trad ← {Slither, Mythril, Echidna, Medusa, Hevm,
Aderyn, Halmos, Manticore, Heimdall, Panoramix}
modules ← {circom, zk-verifier, bridge, oracle, governance,
proxy, bytecode, ai-fuzzing, symbolic, defi, rollup, DAG}
agents ← {Triage, Correlator, AttackChain, Exploit, Remediation, Summary}
▸ Layer 1-2: Parallel tool execution (28 tools)
futures ← ∅
for all t ∈ tools_trad ∪ modules do
futures ← futures ∪ {Async(t.analyze(P))}
outputs ← AwaitAll(futures)
▸ Layer 3: Normalization
findings ← Deduplicate(∪ t.normalize(o) for (t,o) in Zip)
world ← BuildWorld(P, findings)
▸ Layer 4: Sequential LLM refinement
for all a ∈ agents do
findings, world ← a.process(findings, world)
▸ GAEV: Exploit verification for Critical/High
for all f ∈ Filter(findings, severity ∈ {Critical, High}) do
f.exploit ← GAEV(f, world)
return world.generate_report()Generalization (§9)
Code Security
Tools: Semgrep, CodeQL, Bandit, SpotBugs, Coverity
Benefit: Similar FP reduction; business logic understanding
Document Analysis
Tools: OCR (Tesseract, PaddleOCR), Layout analyzers, NER
Benefit: Entity correlation, contradiction detection, summary gen
Medical Diagnostics
Tools: Image classifiers, Lab analyzers, Symptom checkers
Benefit: Differential diagnosis — requires extensive validation