Bodega — state-verified rewards for browser shopping agents

Bodega is a browser-use RL environment where agents shop on a store we own, so rewards come from backend state the agent actually mutated — checked via a private API against our database — not from an LLM judge reading a transcript. This makes rewards near-unhackable for state-mutating tasks and produces a clean training signal.

Why state verification

Every shopping/browser env on the Hub is either read-only navigation scored by an LLM judge or a shopping eval without verifiable rewards. Bodega's differentiator: for cart/checkout tasks, the reward is computed from the store's own database (GET /api/verify/{sid}, bearer-authed), so "did the agent really add 2 blue jackets?" is answered by ground truth, not opinion.

Task tiers

Tier	Task	max_turns	Reward
T1	Lookup (price / rating / spec / stock)	5	answer match
T2	Search + filter superlative	8	answer match
T3	Numeric comparison	8	answer match
T4	Cart build (exact multiset; supersets fail)	10	verify state
T5	Constrained cart (cheaper scores higher)	14	verify state (shaped)
T6	Checkout (items + coupon + shipping exact)	16	verify state

Tasks are procedurally generated with machine-checked rejection rules (unique answers, ambiguity margins, in-stock variants) and split into disjoint sets with an automated dedup check.

Split policy (public dev / private held-out, Zapier-style):

Split	Shipped in package?	Purpose
`train-pool`	yes (public)	training pool
`eval`	yes (public)	public dev/eval set
`test`	no (repo only)	held-out, same distribution — official/uncontaminated numbers
`eval-ood`	no (repo only)	held-out, different catalog — OOD generalization

Only the public splits are bundled into the published environment, so the held-out sets can never leak through the Hub. Official scores are measured locally against test / eval-ood.

Usage

import verifiers as vf
env = vf.load_environment(
    "bodega",
    tier="t1,t2,t3",      # "t1".."t6" | "all" | comma list
    split="eval",          # public: "train-pool" | "eval"  (held-out, repo only: "test" | "eval-ood")
    mode="dom",            # "dom" (default, text model) | "cua" (vision fallback)
    num_examples=-1,
)

Required environment variables

Variable	Used for
`BROWSERBASE_API_KEY`, `BROWSERBASE_PROJECT_ID`	browser provider
`MODEL_API_KEY`	Stagehand DOM executor (pinned `openai/gpt-4o-mini`)
`BODEGA_VERIFY_KEY`	bearer for the store's session + verify API
`BODEGA_STORE_URL`	store base URL

Design guarantees

Infra faults raise, never score 0.0 (D6): dead sessions / 5xx / expired-sid 404 raise vf.Error so prime-rl drops the rollout instead of poisoning the gradient with a fake policy failure.
Deterministic rewards — no LLM judge anywhere.
Pinned Stagehand executor (openai/gpt-4o-mini) — recorded per run; an executor change is treated as an environment change.
Anti-hacking: superset carts fail, multiple orders fail, zero-tool-call rollouts score 0, answers can't be satisfied by echoing the prompt.

License / provenance

Storefront catalog is fully synthetic (seeded generation) — no scraped data.