Built by two founders, for every research lab.
Verifiable Labs is the work of Stelios Zacharioudakis(research & engineering) and Stefanos Eleftheriou (operations). The project grew out of contamination and reproducibility concerns from medical-imaging research at AsklepiosMed.
View on GitHubThe two people behind Verifiable Labs
Stelios Zacharioudakis
Co-founder · Research & EngineeringAthens, GreeceML engineer focused on production ML systems. Sole author of four 2026 manuscripts on LLM inference, formal verification with SMT, trustworthy classification, and dynamic VLBI imaging.
- 4 solo-author manuscripts (2026)
- PyTorch · vLLM · Z3 SMT · GNNs
- AsklepiosMed · 480 doctors · 99.5% uptime
- Head EngineerJun 2022 — PresentPaphos Medical Association · AsklepiosMed
Platform serving 480 doctors across Cyprus. Reduced onboarding 60% via automated digital identity (Stripe → Apple/Google Wallet). Cut deployment cycle from 2 days to 15 minutes; 99.5% uptime over 12 months.
- Full-Stack EngineerJan 2025 — PresentMedihyal Clinic
Clinic booking platform on Next.js 16 + Supabase handling 200+ monthly reservations. Integrated Groq LLM for inventory reorder review, reducing manual review time by 80%.
Stefanos Eleftheriou
Co-founder · OperationsPaphos, CyprusCommunication & Information Studies at the University of Groningen, graduating July 2026. Former member of the Cyprus National Swimming Team, with seventeen years of competitive swimming.
- Cyprus National Swim Team · 17 yrs
- Athletes' Unit · National Guard 2021–22
- WordPress · SEO · Digital Communication
- Marketing & Web AdministrationJul 2025 — PresentDomenica Group
WordPress site administration, SEO, and content strategy across the company's real-estate portfolio. Front- and back-end CMS work, content publishing, and digital brand communication.
- Accounting AssistantJul – Aug 2024ProVision Accountants
Invoice processing and bank reconciliations during a summer placement; financial-records accuracy and compliance work.
Make scientific RL contamination-proof and verifiable, so that every reward signal corresponds to a real-world capability.
Procedural regeneration plus conformal calibration. Closed-form ground truth, distribution-free coverage, fully open source.
Frontier RL has hit a measurement ceiling. Without verifiable rewards, post-training optimizes the proxy instead of the capability.
Why this matters
Most public reinforcement-learning benchmarks were never built to resist contamination. As frontier models scale, the implicit assumption that test instances are unseen quietly breaks. The reward signal stops measuring capability and starts measuring memorization.
Verifiable Labs starts from a different premise. Every problem instance is procedurally regenerated from a physical prior, so the model has, by construction, never seen it. Every reward is bounded by a conformal interval calibrated against a closed-form classical solver. Memorization no longer pays.
None of the primitives are new. Compressed sensing, conformal prediction, and procedural generation predate this work by years. What is new is the combination — and the insistence that scientific RL deserves the same epistemic rigor as medical imaging.
If you’re building post-training pipelines, evaluating frontier models, or shipping AI into healthcare, this is the substrate you can audit. The repo is open. The protocol is peer-reviewed. The environments are live.
Contact
Open to research collaborations, paper discussions, and serious production deployments.