01 / Verified environment
Attested enclaves
Trusted execution environments (TEEs) isolate sensitive computation. Remote attestation lets participants verify the approved software environment before releasing keys.
Intelligence.Confidence.
Confidential AI for clinical trials.
We’re developing a trusted arbiter that computes on unblinded data and releases only pre-authorized signals, helping trial teams act sooner while protecting the blind.
Protected computation
Controlled disclosure
Accountable action
The platform
Compute on the data.
Control what leaves.
Blinding, privacy and sponsor confidentiality protect trial integrity. Tauvera’s architecture brings approved computation inside those boundaries, alongside existing inference engines and trial systems.
02 / Shared key control
Threshold cryptography
Shamir secret sharing splits decryption keys across independent parties. The design requires a threshold of shares, so no single party, including Tauvera, can unlock protected data alone.
03 / Pre-authorized work
AI inside the boundary
Agents analyze data and perform operational tasks within the enclave. A locked specification limits their work and the signals each recipient can receive.
Applications
One arbiter.
Four trial functions.
Our initial research will test whether the same protected core can support safety surveillance, interim readiness, arm-aware resupply and continuous data cleaning while limiting information leakage.
01 / Safety surveillance
Signals for human review
Monitor adverse events by treatment arm inside the enclave. Send designated reviewers a signal when a pre-specified threshold is crossed, without exposing arm-level data to blinded teams.
02 / Investigational-product resupply
Supply informed by actual demand
Use arm-level demand to calculate kit resupply, with controlled instructions sent directly to supply systems or through RTSM/IRT. Evaluate waste, stockouts, and disclosure together.
03 / Interim decision-readiness
Timing set by the data
Track accruing endpoint data by arm inside the enclave as adaptive, platform and event-driven designs make interim timing harder to forecast. Notify the independent statistician when pre-specified information thresholds are met, so analyses run when the evidence is ready rather than on projected dates.
04 / Pre-cut data cleaning
Clean continuously, cut on demand
Run risk-based checks on accruing trial data inside the enclave and route queries to sites without revealing treatment arm. Keep critical data analysis-ready so the DMC can convene on the statistician’s schedule, not after a pre-cut cleaning push.
Beyond these four: the same core extends to confidential eligibility screening and multi-sponsor platform trials.
People retain final authority. Safety, efficacy and trial-continuation decisions remain with authorized human reviewers.
Future applications: confidential eligibility screening and multi-sponsor platform trials.
Trial integrity by design
Authorized signals.
Traceable actions.
The proposed architecture connects existing data and operations systems through the trusted arbiter. Its audit design records each computation and output for authorized oversight, including FDA inspection. The design aligns with ICH E6(R3) risk-based quality management and 21 CFR Part 11 electronic-records requirements.
Proposed trial workflowIntegration and validation ahead
Approved inputs
Trial systems
- EDCClinical and safety data
- RTSM / IRTAssignments and inventory
- SponsorsPermitted data and models
Tauvera
- Attested TEEProtected execution
- Shared key controlNo unilateral decryption
- Restricted agentsPre-authorized tasks and outputs
Permitted outputs
Signals & actions
- Statistician, DMC & safety reviewersThreshold and readiness alerts
- Supply chain / RTSMApproved kit instructions
- Sponsor oversightBlinded operational views
Disclosure is part of the evaluation.Our information-theoretic framework measures excess information leakage relative to a defined trial baseline, including what operational outputs may reveal.
Concept to prototype
Building the evidence.
Completed work includes the trusted-arbiter architecture, formal security specification, leakage framework, and drug-supply simulations demonstrating potential $10-$12 million per-trial savings for Investigational Product/study kits. Next: enclave implementation on commercial confidential-computing infrastructure, then prospective validation.
Meghan Dierks
Zhunping (Justin) Zhang
Collaborate with Tauvera
Protect the blind.
Advance the work.
We're seeking design partners for prospective validation: sponsors, CROs, biostatisticians and trial-technology teams.