AI Governance at the Edge

Four-gate per-node pipeline · F : A → B | P · Regulator-readable evidence

Every agent action becomes a typed transformation that must cross a verifiable edge — before it runs and after it answers. Humans define the contract. GEM² audits at the edge. The LLM operates inside an enforced face.

🦞L0ingress DPI ▶ 🛡L1P-check audit ▶ 🤖Fcontract executor ▶ 🛡L2O-check audit ▶ 🦞L3egress DPI
🦞
Lobster Trap — L0 (ingress) + L3 (egress) "Is this safe?" — pure-Go regex DPI + 5 escalators + LLM intent canonicalize. Catches injections, exfil, credential leak, wire-fraud — on BOTH wires. Vultr DeepSeek-V3.2.
🛡
GEM² Truth Filter — L1 (P-check) + L2 (O-check) "Does this satisfy the contract?" — L1 verifies input against A + P_pre; L2 verifies output against B + P_post. Retrieval-augmented epistemic audit. Gemini 3.1 Pro Preview.
🤖
Contract-Executor (F) — the work itself The agent's actual transformation, bounded by A → B | P. The judge (Gemini) and the worker (Vultr) are never the same LLM — that separation is what makes the verdict trustworthy.

GEM².AI × Veea — TechEX Hackathon 2026 · Track 1: Agent Security & AI Governance
Inseok Seo team lead · Jacobus Wentzel · david@gemsquared.ai

TEST FIELD