DISCOVERY AI / PLANNED
Connect medical evidence and approved research data to accountable AI tasks. Prepare inputs before computation, keep the work within its purpose and assess outputs before they leave.
Explore the required pipelinePlanned pipeline · No live AI execution
BEFORE, DURING AND AFTER AI
Pre-filtering is necessary, but it cannot carry the entire privacy claim. Input preparation, model behaviour and released answers each need their own checks.
Within the originating custodian’s environment, classify the material, minimise the requested information and remove or transform direct and indirect identifiers. Inspect narrative text, attachments, embedded image data and rare combinations.
An approved anonymized artifact, assessment evidence and purpose-specific permission. Unreviewed prompts and identifiable source records are excluded from external AI requests.
Bind the researcher and any agent to a named purpose, approved input artifact, allowed operations, execution environment, recipients and expiry. Separate literature retrieval, inference and model training; approval for one does not authorise all three.
Enforce recipient and tool restrictions, approved locations, retention terms and no reuse for training unless specifically authorised. No arbitrary external-model fallback.
Assess disclosure through answers, small cohorts, repeated queries and model behaviour. Evaluate membership-inference risk at patient level where applicable, including disproportionate exposure of underrepresented groups.
Document the threat model, scientific evaluation and privacy review. Apply appropriate suppression, access restrictions or differential privacy with justified accounting. A failed assessment blocks release.
TWO KINDS OF INTELLIGENCE
Both matter. Their inputs, permissions and validation requirements differ.
Find relevant published literature, retain citations and dates, expose conflicting findings and distinguish a generated summary from its sources. Literature access and licensing need their own arrangements.
Connect evidence to expertiseRun a defined analysis over approved anonymized inputs or within approved local environments. Preserve provenance, review results and evaluate whether a finding is ready to be shared.
Start with the data boundaryCLEAR ANSWERS
Yes. Anonymized data only is the required model for research sharing in Hybrid Discovery. Identifiable patient records stay with the originating custodian; they are excluded from cross-clinic sharing and third-party AI requests. Medical sharing and anonymization services are planned and must pass validation before launch. No patient data is collected or shared by this website.
No. Models and outputs need separate privacy evaluation. Answers can disclose information through small groups, repeated queries or model behaviour even when identifiers were removed from inputs. The required pipeline reviews outputs before release.
No. The required Discovery medical architecture excludes identifiable patient records from third-party AI requests. Approved anonymized material still needs recipient, retention, location and purpose review before any external processing. These controls are planned and require validation before launch.
A proof can establish a precise claim without exposing its underlying witness, under the assumptions of that proof system. It does not anonymize a medical record, establish clinical truth or prevent a model from leaking training information.
ADVANCE DISCOVERY TOGETHER
Build an evidence-led research task with clear input and output decisions.