TonicAI/synthetic_retail_policy_environment

Synthetic Retail Policy Environment

The complete world behind the Fabricate RFT cookbook: a fully synthetic retail customer-support environment with a written store policy, used as a live RL environment to reinforcement-fine-tune a 32B agent (held-out policy application 44% → 67%).

Generated end-to-end by Tonic Fabricate's Data Agent — every customer, order, product, and scenario is synthetic. No real customer data appears anywhere.

What's in it

file	rows	columns
users.csv	300	user_id, first_name, last_name, email, zip, payment_methods (JSON)
products.csv	80	product_id, name, category, final_sale (0/1), variants (JSON map item_id → {options, price, available})
orders.csv	600	order_id, user_id, status, delivered_at, items (JSON), payment_history (JSON)
scenarios.csv	308	scenario_id, user_id, intent, order_id, item_ids, new_item_ids, payment_method_id, reason, difficulty, policy_kind, goal
policy.md	—	the store policy document the agent must apply (30-day return window, final-sale exclusions; reference date 2026-07-15)

Each scenario's gold action is machine-checkable: intent (exchange / return / cancel / modify_items / decline) plus the exact order_id / item_ids / new_item_ids that a correct agent must write — the episode reward in the cookbook is exact-match against these columns. policy_kind tags the policy-stress rows (e.g. window_edge 33, window_expired 32, eligible_subset 90, final_sale_block 33; blank for the 120 standard rows).

How it was used

In the cookbook, this world is served live by Fabricate as mock REST APIs. During reinforcement fine-tuning, every rollout executes the agent's tool calls against these APIs; the reward is a deterministic function of the executed writes vs. the scenario's gold action — no LLM judge. Serving-time writes validate and acknowledge but never mutate, so parallel rollouts share one consistent world.

The trained skill is policy application, not policy memorization: swap the policy document in the prompt (e.g. a 30-day → 14-day return window) and the tuned agent follows the new rules without retraining (violation rate statistically identical to base). Details and experiments are in the cookbook.

Reproduce / extend

You don't need this dump to run the cookbook — its first section contains the Data Agent prompt that regenerates an equivalent environment from scratch in your own Fabricate account. This dataset is the frozen reference world from the published run, for anyone who wants to inspect the data or reproduce numbers exactly.

Sister datasets

• TonicAI/synthetic_clinical_notes
• TonicAI/synthetic_clinical_conversations