About Gerra
Gerra is a data origination company supplying frontier AI labs and institutional researchers. Today, we build professional work collections, full-history software data, and agent environments with trajectories, rewards, and verifiable outcomes for training and evaluation.
Supplying frontier AI labs and leading investment firms.
Current products
Each offer begins with a concrete unit a research team can inspect, evaluate, and extend around the capability it is building.
Source-native internal documents and connected workplace records across engineering, finance, sales, consulting, communications, projects, and company operations. Built for pretraining, post-training, document intelligence, and evaluation.
Editable documents, rendered views, tables, figures, and the native structure that carries the substance of the work.
178 full-history codebases with 15,238 unique commits, plus 2,400 spec-to-repo tasks, 24,000 long-horizon repo tasks, 4,000 exploit-and-patch tasks, and 3,200 browser-graded web builds for pretraining, post-training, RL, distillation, and evaluation.
Cloneable repositories with source, tests, configuration, documentation, branches, merges, and the changes that connect them.
16,000 graded business RL instances and 24,000 computer-use trajectories, plus 400 ML competition runs, 200 paper replications, and 240 post-training runs. Runnable environments, observations, actions, rewards, and final artifacts for long-horizon agents.
The task, source material, tools, starting state, constraints, and resettable execution surface in one package.
Real, de-identified company systems and records, converted into runnable agent environments with expert-defined tasks and private graders.
Company records, communication, actors, policies and business-system state, linked across the workflow.
24,000 browser and desktop trajectories with per-step screenshots, accessibility observations, grounded actions and final-state checks.
The screenshot and accessibility tree available before each action.
Work product and context
A document carries decisions in its tables, figures, edits, and native structure. A software change carries the history that led to it. We preserve those forms, then connect them to the communication, project, and operating records that explain what happened and why.
Explore professional work →Questions, methods, limits
Our public work connects a question to its source material, method, finding, and boundary. The GPU pricing study, for example, found wide market dispersion and published that its tested panel did not produce standalone directional alpha.
Read the research record →Selected lineage
Selected specifications, collection notes and studies from our public research.
A common representation for multimodal robot demonstrations, observations, actions, calibration, coordinate frames, and embodiments.
A four-year, 38-provider study of compute pricing that published the market structure and the limit: no standalone directional alpha in the tested panel.
The construction and verification record for 178 complete repositories and 15,238 unique commits, with a cloneable public sample.
Submitted by Eagle Alpha researcher Mikheil Shengelia and written in collaboration with Gerra; published by Integrity Research with a Michael Mayhew byline.
Research interests
Questions that arise from the material we work with today.
Inspect the work