Ndome is the Gĩkũyũ word for shield — the one a warrior earned by proving themselves. This is that shield for AI agents: it stands between the agent and what it can touch, and it's earned in the open, by proof.

I build the way I believe a safety evaluator has to be built: evidence first, claims second. Over the past year I designed and built Ndome end-to-end as a solo founder — a 7-layer security/QA engine of ~25,000 lines, wrapped in nightly automated testing and 56 graded adversarial attack vectors mapped to OWASP LLM Top-10, MITRE ATT&CK/ATLAS, STRIDE, SLSA, and SOC 2. No team, no institution, no outside funding — just the system, tested continuously and in the open.

What makes the work defensible is the discipline underneath it. Every score carries an explicit C1–C5 certainty grade and a traceable evidence trail. A hard no-laundering rule stops best-case sandbox results from ever inflating the real number, so I publish the honest live-system score even when it's low. I red-team my own claims: on an early blind run, the harness caught a real boundary break I had missed — I withdrew the score, fixed it, and re-verified before crediting anything. I would rather show you the gap than hide it, because that gap is the roadmap.

This rigor isn't new for me. I've applied the same evidence discipline — certainty grading on every claim, strict separation of verified fact from inference — to high-stakes, real-world record-keeping where being wrong carries real consequences. That habit is the core of Ndome: an independent, reproducible, privacy-preserving way to verify what an AI agent actually does, without ever touching the owner's data.