We are researching healthcare foundation models, collaborating agents and better methods of discovery—so people and machines can understand a person’s history, coordinate work and develop better care.
Research stage Study designs and reference code are available. Model experiments and clinical validation remain ahead.
Research multimodal models that connect clinical language, images, device signals and patient experience over time. Study longitudinal context, medical knowledge and uncertainty—including causal questions about what changes outcomes.
Investigate how agents gather evidence, test explanations and coordinate work with care teams and robots. A shared task might involve checking sources, planning follow-up and challenging a proposed approach.
Study whether retained changes to a research procedure help later experiments find better models or methods with comparable resources. Learning new model weights and improving the research process are separate questions.
We’re developing an AI agent that carries authorized context across tasks, turns reviewed experience into reusable skills, and helps care teams investigate evidence and coordinate work.
VANTHEON Agent is the system for doing that work. THEON is the foundation-model family and research program intended to advance its intelligence.
Help shape VANTHEON AgentA corrected observation can affect a care decision, a follow-up plan and later research. Using synthetic care episodes, this study asks whether an agent can identify what needs reassessment while preserving uncertainty and unrelated evidence.
It tests evidence handling, not clinical benefit. Model learning, patient outcomes and physical safety need their own studies.
Read the study designAuthorized care experience can inform THEON; independently evaluated findings can inform Codex and care tools. The proposed Network Treasury supports research separately from the Pool’s care funds.
Learn, verify, settleInspect the study designs, reference code and proposed evaluation methods.
A separate proposed study of model adaptation using synthetic care records.
Study specifications and deterministic replays of research procedures—not model-performance results.
Same starting model, data, tools and total budget.
Protected tests.
All attempts and resources counted.
Register the task and budget; compare against fixed and single-agent baselines. Protect the final assessment and count failed attempts, human assistance, subgroup performance and total resource use.
Model weights and research procedures are versioned separately. Independent review determines the supported scope, with monitoring and a way to fall back. Clinical and financial decisions retain their own authority.
Care consent does not automatically authorize training. Research use, contributor rights and withdrawal limits need their own terms.
Sovereign operation means the ability to inspect, export, restore and replace models, data and systems, with control over keys and visible external dependencies. Corrections trigger review of affected datasets, models and uses.
