Themis

The problem

Legal and judicial data management is labor-intensive and delay-prone. The dominant driver of time-to-court-martial is the sheer volume of digital evidence — phone extractions, message threads, documents, images, audio — that judge advocates must read, evaluate, and synthesize against the elements of each candidate charge, a mapping still built largely by hand. It is not primarily a model-quality problem; senior-attorney review already catches errors as it always has. It is an evidence-processing and workflow problem, made harder by constraints unique to legal practice: attorney-client privilege, work-product protection, a defensible discovery posture, and security boundaries where commercial frontier-model access is limited.

Our approach

• •Privilege-aware at ingest – protected material is tagged and segregated the moment it enters, with role-based access and a complete audit log of what was ingested, what was produced, and who saw it.
• •Structured evidence synthesis – multi-modal evidence is indexed, deduplicated, and mapped to the elements of candidate offenses — a machine-assisted version of the charge-elements chart attorneys already build, with every mapping traceable to its source items.
• •Bounded by construction – classifiers are wrapped in conformal selective prediction, so outputs carry distribution-free, finite-sample confidence and the system explicitly abstains and routes to human review when it does not know.
• •Process automation where risk is lowest – the routine clerical layer of the same workflow (file organization, form population, deadline and docket tracking) is automated for immediate time savings, with human judgment preserved at every consequential step.

What makes it different

Off-the-shelf generative legal assistants (Harvey, CoCounsel, Lexis+ AI) are built on commercial frontier-model APIs and cloud delivery, which places them outside the deployment boundary available to many legal offices — and their assurance rests on benchmarks and review, without formal guarantees. Themis is the combination that isn’t otherwise available: element-schema-native synthesis, conformal selective prediction with explicit abstention, and privilege-by-design architecture that runs on-premises without frontier-model dependence. The design is retrieval and structured synthesis, not generative-first — lower-risk, more auditable, and explainable by construction rather than by a post-hoc heatmap.

Status

Themis is an active research and development effort — a U.S. Air Force SBIR Phase I feasibility study (topic DAF26BZ03-NV018) for military justice, executed entirely on unclassified public and synthetic data. The immediate customer is the Air Force District of Washington Judge Advocate (AFDW JA) and the broader AF/JA enterprise; the clearest enterprise scaling path runs through the Offices of Special Trial Counsel (OSTC), whose independent, covered-offense caseloads are the most digital-evidence-intensive in military justice. It builds on the team’s peer-reviewed conformal-prediction research and delivered, calibration-first multi-modal ML systems, and is grounded in practice through an advisory judge advocate.