RESEARCH

Research-grade data most labs will never see.

A decade of real investigative operations — reasoning chains, human–machine teaming telemetry, doctrine structure — available to researchers on our platform. We publish, we collaborate, and we are actively looking for researchers to work with us.

Apply for research access
FEATURED PAPERS

Current working papers.

UPDATED WEEKLY
RT-2026-01 · DRAFT
Doctrine as Substrate: Grounding Model Reasoning in Investigative Methodology

Encoding schoolhouse intelligence doctrine as machine-usable structure, and measuring its effect on long-horizon reasoning quality against undoctrined baselines.

STATUS: INTERNAL DRAFT · SEEKING COLLABORATORS
RT-2026-02 · DRAFT
Divergence Curves: Where Human and Model Judgement Separate in Live Casework

A decade of teaming telemetry reveals systematic, predictable divergence between operator and model decisions — and where override patterns signal capability gaps.

STATUS: DATA ANALYSIS · SEEKING COLLABORATORS
RT-2026-03 · DRAFT
Grounded Synthetic Data: Generating Training Corpora from Provenanced Casework

Synthetic expansion of real investigative data that preserves evidentiary structure and provenance — scaling a private corpus without inheriting the failure modes of pure synthesis.

STATUS: METHODOLOGY DRAFT · SEEKING COLLABORATORS
ACTIVE ANGLES

Five research directions, open for collaboration.

A-01
Doctrine-grounded reasoning
How encoded investigative doctrine changes model behavior on long-horizon analytical tasks.
A-02
Human–machine teaming dynamics
Behavioral and representational divergence between operators and models across a decade of shared casework.
A-03
Layout-aware retrieval over evidence
Graph-structured retrieval architectures that preserve document layout and evidentiary relationships.
A-04
Grounded synthetic data
Provenance-preserving synthetic expansion of operational corpora for training and evaluation.
A-05
Corroboration under uncertainty
Formalizing single-source corroboration doctrine for machine consumption — when should a system trust one witness?