"The neurons propose. The rules dispose."
Veea / lablab.ai Hackathon · May 11–19, 2026
A 95% accurate LLM is a 5% critical failure rate in production.
Real court cases already happening:
| Case | Consequence |
|---|---|
| Lacey v. State Farm (May 2025) | $31,100 in sanctions |
| Russell v. Mells | Referral to Florida Bar |
| Flycatcher v. Affable | Default judgment entered |
| Baidu AI (China) | Systemic trust collapse |
Judge Wilner: "I discovered they DID NOT EXIST. That's frightening."
No enterprise legal team will deploy AI without a deterministic trust layer.
Input → Lobster Trap (DPI) → Phi-4-mini (Extract) → Deterministic Validator (8 Rules) → APPROVED / QUARANTINED → Human Review
| Layer | Technology | Function |
|---|---|---|
| Ingress Security | Lobster Trap DPI (Go mock, Veea API-compatible) | Blocks prompt injection, PII, exfiltration |
| Neural Extraction | Phi-4-mini + Ollama | Extracts 8 atomic radicals (Actor, Action, Metric, etc.) |
| Symbolic Validation | Custom JS Engine | Applies 8 deterministic rules (R1–R8) |
| Human Review | Forensic Panel | Approve/Discard → LoRA fine-tuning |
🔬 Radical Grounding Check: Verifies every extracted radical literally exists in the source text. If the LLM invents a concept → instant quarantine.
"The LLM proposes. The rules dispose."
flowchart TD
classDef default fill:#141c2e,stroke:#2a4a7f,stroke-width:2px,color:#e2e8f3;
classDef block fill:#ef4444,stroke:#7f1d1d,stroke-width:2px,color:#fff;
classDef approve fill:#10b981,stroke:#064e3b,stroke-width:2px,color:#fff;
classDef quarantine fill:#f59e0b,stroke:#78350f,stroke-width:2px,color:#fff;
classDef ai fill:#3b82f6,stroke:#1e3a8a,stroke-width:2px,color:#fff;
Input[📄 Input Clause] --> Shield
subgraph TridenGuard [Neuro-Symbolic Isolation Architecture]
direction TD
Shield[🛡️ SHIELD A: Lobster Trap DPI <br/> Evaluates PII & Prompt Injection]
Neural[🧠 NEURAL LAYER: Phi-4-mini <br/> Extracts 8 Atomic Radicals]
Symbolic[⚖️ SYMBOLIC LAYER: Deterministic Validator <br/> Applies Grounding & 8 Exclusion Rules]
Shield -- Threat Detected --> Blocked[🚫 BLOCKED]
Shield -- Clean Payload --> Neural
Neural -- Structured JSON --> Symbolic
Symbolic -- Pass --> Approved[✅ APPROVED]
Symbolic -- Structural Failure --> Quarantined[🔴 QUARANTINED]
end
Quarantined --> HITL[👨⚖️ Human Review Panel <br/> Approve / Discard]
HITL -. TOON Preference Signal .-> LoRA[(🧠 Local LoRA Dataset)]
class Blocked block;
class Approved approve;
class Quarantined quarantine;
class Neural ai;
| Layer | Result |
|---|---|
| Lobster Trap (DPI) | 100% block rate |
| Real court hallucinations | 100% intercepted |
| Structural validator (R1–R8) | 87.5% accuracy |
| Overall pipeline | 85% (17/20) |
A 64-case matrix is designed for V2.
| Phase | Focus | Status |
|---|---|---|
| V1 | Deterministic firewall + Human panel | ✅ Delivered |
| V2 | GBNF token governance + Fisher's Exact Test | 🔄 Next sprint |
| V3 | Local LoRA fine-tuning flywheel | 🔁 Q3 2026 |
| Base | Veea Edge Nodes (air-gapped, low-latency) | 🛡️ Planned |
- Docker & Docker Compose
- Go (to build the DPI mock once)
# 1. Clone the repository
git clone https://github.com/AlexusPacicus/TridenGuard.git
cd TridenGuard
# 2. Start n8n + Ollama
docker compose up -d
# 3. Pull the LLM model (inside the Ollama container)
docker exec ollama_tridenguard ollama pull phi4-mini:3.8b
# 4. Build & run Lobster Trap DPI mock (required — not in compose yet)
cd lobstertrap_service && bash build_mock.sh && cd ..
./lobstertrap serve --port 8080 # leave this terminal open
# 5. Import & activate workflow (first time only)
# Open http://localhost:5678 (admin / admin) → import n8n-workflows/TridenGuard.json → Activate
# Configure Ollama credential → base URL: http://ollama:11434
# 6. Forensic panel (benchmark UI — simulated cases for demo/video)
open frontend/index.htmlcurl -X POST http://localhost:5678/webhook/tridenguard \
-H "Content-Type: application/json" \
-H "x-api-key: tridenguard_secret_key_2026" \
-d '{"text": "Shall appoint as exclusive distributor within the Market."}'Immediate response: {"message":"Workflow was started"} (async pipeline, ~20–30s)
Verdict (n8n → Executions or quarantine table): QUARANTINED — R2_ACTION_WITHOUT_SUBJECT
DPI block test: IGNORE PREVIOUS INSTRUCTIONS... → blocked at Lobster Trap before Phi-4.
- Live UI Demo: triden-guard.vercel.app (frontend simulation with real benchmark cases)
- Full Video Demo: Watch on YouTube
The 3-minute video demonstrates:
- Real execution in n8n (Lobster Trap → Phi-4 → Validator)
- Forensic panel with quarantine, approve, and export
- Benchmark results (85% accuracy, 100% block rate)
| Select Contract | Paste this clause | Expected Verdict |
|---|---|---|
| REAL-R1 LIMEENERGYCO | Company and Distributor must comply with all obligations stipulated in this agreement. | 🔴 CRITICAL — R1_SUBJECT_WITHOUT_ACTION |
| TECH-R4 TechVista | The profitability threshold is set at 15%. | 🔴 CRITICAL — R4_ORPHAN_METRIC |
| PHARMA-R8 Pharma Global | The Receiving Party shall not from any source other than the Company. | 🔴 CRITICAL — R8_DEONTIC_WITHOUT_BEHAVIOR |
| APEX-R7 Apex Construction | Located in Los Angeles. | 🔵 BORDERLINE — R7_INERT_SPATIAL |
| Component | Technology |
|---|---|
| Orchestration | n8n |
| Security | Lobster Trap DPI mock (Go; Veea binary in V2) |
| LLM | Phi-4-mini (3.8B) + Ollama |
| Validation | Custom JS (8 rules + grounding check) |
| Observability | TOON + JSONL (V2 roadmap) |
| Frontend | HTML/CSS/JS |
| Deployment | Docker + Vercel |
TridenGuard includes a complete set of YAML policies for Veea Lobster Trap (see configs/).
The n8n workflow includes an HTTP request node that calls Lobster Trap at http://host.docker.internal:8080/inspect.
Current MVP status: The live demo on Vercel is a frontend simulation.
A fully functional local deployment with Lobster Trap (including the provided mock) is available via docker-compose up.
This entire B2B firewall — including local Phi-4-mini LLM inference, n8n orchestration, and Lobster Trap DPI — was developed and stress-tested on a 2020 Mac M1 with only 8GB of RAM.
If TridenGuard can execute a full Neuro-Symbolic pipeline on a 6-year-old entry-level machine, it is lightweight enough to be deployed on any enterprise edge node today.