Expert-graded legal-reasoning data and training environments.
Amo Ventures builds evaluation sets and reinforcement-learning environments that test whether models can reason about legal standards the way a practitioner does, with gold answers and grading you can inspect.
Products
Two complementary products built around one question: would a reasonable consumer be misled by this food-label claim? Every item is built on made-up products, never on real brands.
Label-Deception Bench
An expert-graded dataset of food-label claims, judged under a consumer-protection reasonable-consumer standard. The model reads label text and decides whether a claim is viable, names the legal theory, and explains its reasoning.
- Gold verdict and theory for every case
- Written expert rationale
- Criterion-level scoring rubric per item
- Covers both actionable and non-actionable claim patterns
- Delivered as JSONL
Label-Deception Gym
An interactive RL environment of 640+ tasks with a deterministic grader. Agents use tools to inspect a simulated store aisle, then submit a decision: the deceptive claim and the ingredient that falsifies it, or a reason the claim does not hold up.
- Agentic, tool-use tasks with JSON tool schemas
- Deterministic grader with partial credit and an efficiency penalty
- Easy, medium, and hard difficulty tiers
- Python environment, grader, and test suite included
- Seeded, reproducible generation with validation
Approach
Legal judgment is hard to fake and easy to grade badly. We design for signal you can trust.
Expert-authored
Cases are written and graded by a consumer-protection practitioner, applying one documented standard.
Made-up products
Every label is a fictional composite. Built only from public regulations, agency guidance, and public claim patterns.
Verifiable reward
Gym grading is deterministic, so a reward means the task was actually solved, not guessed.
Hard negatives
Not-viable cases teach models when a claim that sounds suspicious does not meet the standard.
Licensing
Flexible terms for labs evaluating, training, or benchmarking legal reasoning.
Exclusive dataset license
Exclusive rights to the Bench, limited by time or field, at a premium.
Non-exclusive environment license
Use the Gym for training and evaluation alongside other licensees.
Private held-out evaluation
We run your model against cases never released, so results stay uncontaminated.
Quarterly updates
Optional refreshes with new cases and tasks delivered each quarter.
Custom lanes, jurisdictions, and task counts can be built to spec. Sample items are available under NDA for evaluation.
Get in touch
Interested in a sample, a held-out evaluation, or a license? Reach out to Dan Harvath, founder.
dan@amoventuresllc.com