RL environments · Training data · EVM security

RL environments
and data for
EVM security.

We build executable smart-contract tasks and agent trajectories for post-training teams. Start with a paid pilot scoped to your model and the capabilities you want to improve.

01 / A task becomes training data
Illustrative workflowDefine a security task. Capture tool use. Verify the outcome. Export the trajectory.
Built around your model’s capability gaps
PatchExploitLearn

01 / Research talk

How to benchmaxx
on any benchmark.

Presented to investors and researchers, including researchers from Microsoft Research and IISc.

Watch on YouTube

02 / The offering

Executable environments.
Inspectable training data.

Choose environments, trajectory data, or a pilot that connects both.

RL environments

EVM RL environments

Custom patch and exploit tasks with executable checks and reference solutions. Set protocol coverage and difficulty around your model.

  • Runnable state, so your team can reproduce the task
  • Verifiers and reference solutions to check behavior
  • Difficulty requirements tied to your target model
Scope an environment pilot
Training data

Verified agent trajectories

Tool use linked to task outcomes. Inspect successful runs, failures, and recovery examples for your post-training pipeline.

  • Tool calls, observations, and code edits in context
  • Verification and reward metadata alongside the run
  • Provenance and formats scoped to your pipeline
Scope a data pilot
Delivery, scoped to your stack

Containerized tasks · JSONL / Parquet trajectories · Formats and license terms agreed per pilot.

03 / How we work

Start with a capability gap.
Build a focused pilot.

01

Scope the work.

Define the target model, task families, and training requirements.

02

Build and verify.

Agree on reference solutions, verification, provenance, and difficulty checks.

03

Evaluate the pilot.

Inspect the artifacts and assess performance on held-out tasks.

We agree on deliverables and acceptance criteria before the pilot begins.

04 / Research & technical writing

Inside the work.

Model behavior, training data, and agent tooling.

All research & writing

Let’s build your pilot

What should your
agent learn next?

Tell us about your model, the failures you’re seeing,
and the environments or data you need.

Discuss a pilot