Long-horizon execution
Multi-step work across documents, databases, spreadsheets and messaging systems.
For frontier AI teams
Training infrastructure for systems that must do real work.
Start with a capability gap. Sofitra finds the operating data behind it, reconstructs the world in which the work happens, and delivers tasks, trajectories and rewards that can move model performance.
The mandate
Static datasets show a model the answer. Real environments require it to understand state, choose tools, recover from mistakes and produce an outcome that survives verification.
We build those environments from the operating histories of real companies, so the difficulty comes from the work itself, not a synthetic proxy.
Capability gaps
Built around what your model cannot yet do.
We work backwards from failure: the economic task, the missing context, the required tools, the decision points and the reward signal.
Source-grounded judgment
Claims and decisions tied to authoritative evidence rather than plausible text.
Error detection and repair
Tasks that require finding broken work, diagnosing the cause and recovering safely.
Hidden-rule compliance
Permissions, agreements, policies and edge-case restrictions that cannot be ignored.
Engagement model
From research question to runnable environment.
A narrow pilot can validate task difficulty and verifier quality before scaling into a full training curriculum.
You define
01
02
03
04
Target capability
The work your model must learn or the benchmark it must survive.
Observed failures
Examples of current errors, shortcuts, saturation or reward hacking.
Runtime constraints
Models, tools, container standards, latency and security requirements.
Training objective
RL, supervised trajectories, evals, reward research or a combination.
Sofitra delivers
01
02
03
04
Runnable environment
Resettable state, tools, data, permissions and task interfaces.
Task curriculum
Difficulty bands, hidden holdouts, adversarial cases and refresh plan.
Trajectories and repairs
Expert demonstrations, rejected paths, corrections and recovery traces.
Verifier and evidence pack
Reward code, rubrics, lineage, methodology and evaluation report.
Delivery anatomy
Everything needed to train, inspect and iterate.
Sofitra environments are model-agnostic and designed to fit the lab’s existing orchestration, evaluation and data-governance stack.
Container or API runtime
Deterministic resets, seeded state, controlled tool access, event logs and reproducible execution.
Structured task manifests
Initial state, success criteria, hidden constraints, available tools, metadata and split assignment.
Action-level trajectories
Tool calls, intermediate artifacts, corrections, approvals and outcome-linked reasoning evidence.
Composable verifier stack
Outcome checks, process checks, expert rubrics, hidden tests and reward-component diagnostics.
Pilot structure
Prove signal before scaling supply.
A focused engagement tests whether the environment exposes a meaningful capability gap and produces stable reward before a broader programme begins.
Scope
Define capability, model baseline, tools and acceptance criteria.
Source
Select the data universe and complete rights and privacy review.
Build
Reconstruct state, package tools and author the first task family.
Verify
Implement checks, hidden cases and expert adjudication rules.
Baseline
Run selected models, analyse failures and calibrate difficulty.
Deliver
Ship runtime, tasks, trajectories, verifier and evidence pack.
Scale
Expand domains, cases, horizons and refreshes from new failures.
Engagement
Bring us the failure worth fixing.
Share a target capability and we will propose the data, environment and verifier architecture around it.