STEADYWRK AI LAB · Aqaba, Jordan
مبنيٌّ في الإنتاج. ينشر حين يُثبَت.
The STEADYWRK AI Lab does not research systems that are not running. Every publication in the lab index describes something operating live from Aqaba — measured, stress-tested, and published with a verifiable ground-truth source. No claims without proof. No announcements before production.
Research lines
Production evaluation of AI routing decisions at 340ms median latency. The lab measures, retests, and publishes what the system actually achieves — not what it claims. Eight evals on a rolling 30-day window, served from a public no-auth endpoint.
12 AI agents coordinated by a multi-layer control plane. The lab researches how agents route decisions, escalate under uncertainty, and maintain an audit trail on every action. System card published openly.
How do you measure an AI system honestly? The lab publishes its methodology: ground-truth definitions, sample-size gates, and retraction protocols. A number is not published until the lab can stand behind the ground-truth source.
OpenAPI 3.1 spec, TypeScript SDK, Python SDK — the same libraries the control plane uses internally, exposed for external builders. The lab maintains the spec and the SDKs as first-class research artifacts, not afterthoughts.
EU AI Act obligations, AI accountability frameworks, and data-handling standards for agentic systems. The lab publishes its policy position openly because the obligations apply to any system deployed for real decisions.
Lab outputs — live
Lab doctrine
01
No publication describes a system that is not running. If it is not live, the lab does not write about it.
02
Every metric links to its ground-truth source: a JSON endpoint, an audit log, or an operational timestamp. "We measured it" is not a source.
03
When a number is retired, the retraction is logged at /proof/attestation alongside what replaced it. Integrity is cumulative.
Start with what the lab has already proven. Everything else follows from there.
— V, Aqaba