Sofitra preserves the structure that makes real work difficult (tools, state, permissions, dependencies, exceptions and expert review) and packages it into a resettable system for training and evaluation.
An environment is not a folder of files with a question attached.
It is a living state machine: the model must inspect evidence, operate tools, make changes, obey hidden constraints, recover from errors and leave the world in a verifiably correct state.
Every task is connected to an authorised data source, an executable environment state and a measurable reward specification.
A privacy-safe, internally coherent information universe reconstructed from real company artifacts and histories.
Tools, permissions, starting state, transitions and hidden dependencies required to reproduce the workflow.
Tasks, trajectories, difficulty distributions, verifiers and model-failure feedback loops.
The environment exposes only the information and tools available to the professional at that moment. The verifier sees more.
The starting condition, hidden facts and consequences that make each run meaningful.
The applications and operations the model can use to inspect, edit and submit work.
The condition the model must create, not merely the prose it must produce.
Instead of paraphrasing one problem repeatedly, Sofitra uses historical cases, corrections and failure patterns to vary the state itself.
Representative end-to-end tasks that establish the core skill and expected professional standard.
Rare but consequential situations recovered from actual review and escalation histories.
Partially completed or corrupted work that demands diagnosis rather than generation from scratch.
Cases with incomplete evidence, competing constraints or multiple defensible actions.
Tasks designed to expose memorisation, template following, superficial citation and reward gaming.
New distributions generated directly from observed model weaknesses during evaluation and training.
A single scalar score can hide the reason a run failed. Sofitra exposes reward components and evidence at the level researchers need.
Formulas, reconciliations, database state, file structure, numerical tolerances and policy conditions.
Citation resolution, source authority, claim coverage, temporal validity and contradiction checks.
Prohibited actions, required review steps, tool-use sequence, approvals and state-changing behaviour.
Domain-specific quality, materiality, escalation, communication and decision standards.
We test whether a model can earn reward without genuinely completing the task, and harden the environment against that path.
Unseen data, rules and state checks prevent overfitting to the visible specification.
Critical outcomes are validated through more than one data source or calculation path.
Inputs and environment state are mutated to test whether the strategy generalises.
State transitions reveal destructive shortcuts, fabricated evidence and prohibited operations.
The same environment specification can be packaged for training, private evaluation, model comparison or internal research infrastructure.
Versioned images with reset hooks, seeded state and structured event output.
Model-agnostic action schemas compatible with existing agent harnesses.
Structured metadata, splits, trajectories, artifacts and verifier diagnostics.
Lineage, transformation, permitted use, environment cards and version history.
Start with a task family, a model failure or a real workflow that is not represented in public data.