Tempus AI Medical Intelligence

Tempus Reality

Tempus AI runs commercial molecular sequencing at scale through its xT, xR, and xF assay families, generating one of the largest connected libraries of solid-tumor sequencing data paired with longitudinal clinical and outcome records in oncology. Tempus Next layers algorithmic care pathways on top of that substrate, surfacing eligible therapies, trial matches, and care-gap notifications back to ordering oncologists. Adjacent product lines extend the same multi-modal data fabric into cardiology, neuropsychiatry, and radiology, with imaging, electrocardiogram waveforms, and pathology slides feeding the same connected patient model that genomic data feeds.

The institutional position is unusual: Tempus simultaneously operates as a CLIA-certified diagnostics laboratory, a clinical-grade software vendor, a contract research organization for pharmaceutical partners, and a real-world evidence supplier feeding regulatory submissions. Each of these roles imposes a distinct evidentiary regime — laboratory developed test validation, software-as-medical-device life-cycle controls, clinical trial Good Clinical Practice, and regulatory-grade real-world data quality frameworks — and each regime ultimately resolves into the same operational question: which examples were used to train this model, who was authorized to contribute them, and under what consent did the underlying patient data enter the system.

Path Forward

The trajectory pulling Tempus toward training-governance substrate is the convergence of three regulatory lines that were previously parallel. The FDA's evolving framework for AI-enabled medical devices, including Predetermined Change Control Plan authorization, contemplates models that continue to learn from post-market data and require demonstrable training-time governance to maintain authorized status. Real-world evidence guidance for regulatory submissions increasingly demands provenance-grade documentation of how each clinical record entered the analytic substrate. Patient-data consent regimes — HIPAA, state genetic privacy statutes, and contractual research-use limits negotiated with each contributing health system — define a per-example permission surface that the training pipeline must respect at gradient-application granularity rather than at dataset-aggregate granularity.

The forward operational reality is that an oncology model recommending a targeted therapy, or a cardiology model flagging structural heart disease, will be expected to answer questions about the specific patient cohorts, the specific health-system contributors, and the specific consent envelopes that shaped its parameters. A platform that can answer those questions at training-step granularity — rather than relying on aggregate dataset cards — gains regulatory durability. A platform that cannot, retrains under audit pressure.

The cost asymmetry compounds the trajectory. A retrained-under-audit cycle in a regulated medical-AI platform consumes engineering time, regulatory bandwidth, and clinical-customer trust simultaneously, and is not recoverable through documentation after the fact. The architectural decision to make training-step provenance a first-class property is therefore not a compliance cost but a deferred-rework hedge whose value compounds across every model generation, every regulatory filing, and every partnership the platform takes on.

Training-Governance Substrate

Depth-selective training governance treats the gradient itself as the unit of governance, not the dataset and not the checkpoint. Each training example carries a contribution credential identifying the institution that supplied it, the consent scope that authorized it, and the clinical or molecular evidence class it represents. The training loop reads those credentials and routes the resulting gradient to model layers whose authorization envelope admits that class of evidence. Genomic-evidence gradients flow into representational layers credentialed for sequencing-derived inference; outcome-evidence gradients flow into prediction heads credentialed for outcome modeling; consent-restricted gradients are excluded from layers that serve uses outside the consent scope.

Credentialed contribution attestation supplies the cryptographic backbone. Each contributing site signs a manifest binding example identifiers to consent class, IRB authorization, and contributor identity; the training operator counter-signs the resulting gradient routing decision. The consequence is a per-step, per-layer audit record that survives federation: when Tempus composes training contributions from multiple health-system partners, pharmaceutical collaborators, or external real-world data vendors, the federated training run produces a single attestation chain rather than a set of detached dataset cards. FDA-relevant audit, patient-consent verification, and pharma-partner data-use accounting all read from the same substrate.

Differentiation

Several adjacent architectures address parts of the problem. Federated-learning frameworks keep raw data inside contributor institutions and exchange gradients across the boundary, which is necessary but not sufficient: a federated gradient is still consent-opaque unless the contribution credential travels with it, and is still layer-undifferentiated unless the receiving model routes it by evidence class. Differential-privacy noise injection bounds re-identification risk on aggregate gradients but does not produce per-contribution attestation. Dataset cards and model cards document training data at aggregate granularity but cannot answer audit questions at gradient or layer resolution. Depth-selective training governance with credentialed contribution attestation is the substrate that composes with each of these techniques rather than replacing them, and produces the training-step-resolution audit record that the others do not.

The differentiation that matters commercially is that Tempus's pharmaceutical and health-system partners increasingly require their contributions to be governed not as a pooled training corpus but as a credentialed contribution whose use respects the negotiated scope. A platform that can demonstrate that property at training time wins partnerships that a platform relying on aggregate documentation cannot.

Tempus Position

Tempus already operates the data-integration substrate, the clinical contributor network, and the regulatory engagement that a credentialed-training architecture presupposes. What it does not yet operate, in the public technical record, is an architectural primitive that binds gradient application to contributor credential and consent scope at training-step resolution. Adopting depth-selective training governance with credentialed contribution attestation converts existing data-governance commitments into a training-time architectural property rather than a downstream documentation artifact, and gives Tempus a regulatory posture aligned with where AI-enabled medical-device authorization is heading rather than with where it has been.