Project summary

I'm Emmanuel King, a 26-year-old self-taught researcher in Lagos. Four months ago I started building Lex Aureon because I couldn't find an AI safety layer that provided any mathematical guarantee about behavior at inference time. Everything I found was either training-time alignment (RLHF, Constitutional AI) or reactive prompt filtering. No one was treating runtime safety as a control problem.

So I built one.

Lex Aureon is an open-source constitutional governance layer that sits above any LLM or agent. It models system state as a point on the probability simplex x = (C, R, S) Continuity, Reciprocity, Sovereignty, with a stability margin M = min(C, R, S) and a hard constitutional floor. When a prompt pushes the state toward collapse, a Control Barrier Function governor projects it back into the viable region. Every governed turn generates a SHA-256 cryptographic receipt recording the exact constitutional state, the input hash, and the output hash persisted append-only and publicly auditable.

The system is live right now at

lexaureon.com.

The code is fully open-source https://github.com/omomehinemmanuel5-boop/LEX-Aureon/tree/main

The benchmarks are public. You can hit the API and get a governed response with a constitutional state vector and a receipt ID in seconds.

Core Innovation

Instead of reactive prompt-based or classifier guardrails, LEX-Aureon uses a formal dynamical systems approach:

•  C + R + S = 1 simplex with min(C,R,S) as the stability margin

•  Control Barrier Functions and Lyapunov stability analysis for real-time safety enforcement

•  Self-referential embeddings and multi-agent separation of powers

•  Adaptive constitutional temperature and brittleness metric

Every run generates cryptographically signed (SHA-256) audit receipts for full verifiability.

Results So Far

• 7 benchmark improvements live at lexaureon.com/benchmarks (bare vs. governed).

• 100% blocking of destructive tool calls in the agentic layer.

• 110,000+ turns with cryptographic audit receipts.

• Methodology updated: earlier 0.0% ASR (self-judge) now replaced with grounded judges.

What are this project's goals? How will you achieve them?

Goal: Get the mathematical claims independently validated and the framework into the hands of developers who can use it.

How:

1. Run official classifiers. Use the fine-tuned HarmBench/JailbreakBench judges (not self-judges) to establish citable, reproducible numbers.

2. Red-team the agentic layer. Hire an independent security engineer to attempt bypasses and document failure modes.

3. Remove adoption friction. Build LangChain and LlamaIndex SDKs so developers can add governance with one line of code.

4. Submit to peer review. Write a workshop paper for NeurIPS/ICML presenting the framework and live validation results.

Independent validation. Developer-ready tooling. Peer-reviewed credibility. This is how Lex Aureon moves from a solo project to a standard.

How will this funding be used?

The $200k will be used as follows to move LEX-Aureon from a strong solo project to a rigorously validated and scalable governance layer:

•  $60,000 — Independent red-teaming, third-party audits, and formal verification of the 0% ASR claims and mathematical guarantees (Lyapunov, Control Barrier Functions).

•  $70,000 — Hire 2–3 part-time contractors (developers, technical writer, security engineer) for 6–12 months.

•  $35,000 — Compute resources, proxy optimization, infrastructure scaling, and higher throughput.

•  $15,000 — Documentation, SDKs, and integrations (LangChain, LlamaIndex, CrewAI, etc.).

•  $10,000 — Research, expanded benchmarks, papers, and conference submissions.

•  $10,000 — Public dashboard, transparency tools, operations, and contingency.

Expected Outcomes: Independent validation reports, production-grade reliability, easier developer adoption, faster feature development, and stronger positioning in AI safety.

This is high-leverage funding — the full live system (proxy, 10-agent pipeline, cryptographic auditing, and website) was already bootstrapped solo from Lagos.

Who is on your team? What's your track record on similar projects?

Current: Solo. I designed and built the entire stack—mathematical framework, TypeScript kernel, 10-agent pipeline, cryptographic receipt system, live infrastructure, and CI/CD. No co-founders, no employees, no institutional affiliation.

What are the most likely causes and outcomes if this project fails?

Knowledge Contribution: If the implementation reaches its limit, the research methodology, constitutional logs, and benchmarking data serve as a high-value contribution to AI safety literature and open-source intelligence frameworks.

Iterative Evolution: Failure is leveraged as a "test-to-failure" model, providing the necessary data to refine the Sovereign Intelligence Architecture (SIA) for more robust future iterations.

Technical Asset Retention: The codebase remains a modular foundation, ensuring that intellectual capital is preserved and transferable to future research domains.

Aureonics science is more than software products, but a research framework.

How much money have you raised in the last 12 months, and from where?

$0