earakely-scale/FederalRegisterEnv-REA

federal-register

federal-register

An AgentEnv bundle with its recorded runs, published with agentenv-hf.

Tables

• federal-register_tasks: one row per task of the bundle, with its prompt, envs, verifiers and steps.
• federal-register_episodes: one row per recorded run, with its status, model, reward and scores, the prompt and response, and the agent's transcript as chat messages when its trajectory format is one agentenv-hf reads.

raw/federal-register.jsonl holds each run's record and native trajectory, and bundles/federal-register/ the bundle itself.

Run it

pip install agentenv-framework huggingface_hub
hf download earakely-scale/FederalRegisterEnv-REA --repo-type dataset --local-dir agentenv-dataset
agent-env run agentenv-dataset/bundles/federal-register

Federal Register API on AgentEnv, from a REA capture

What this is. A replay environment of Federal Register API v1 (U.S. public domain), built from one recorded browser session against the public API at https://www.federalregister.gov on 2026-10-19. The responses are real public-domain records published by the Office of the Federal Register, frozen as they were served then. REA projected the capture into an environment manifest (10 tools), and rea environment-export wrote the bundle you can run from this dataset. Every tool call replays a retained record from the capture, cites the record and the Evidence it came from, and nothing executes Federal Register API. Tasks are graded deterministically, with no judge.

What's in it

• 10 tools (10 GET) over 52 recorded exchanges on https://www.federalregister.gov; no recorded group excluded.
• Path parameters: 3 sibling-variance. REA's four format rules (numeric, uuid, hex, iso-date) applied unconditionally; the three opt-in rules (after-uid-literal, opaque-token, sibling-variance) were requested and are labelled per parameter. With the four format rules alone the manifest has 45 tools.
• Unknowns the manifest states explicitly: authorization model, response semantics beyond recorded exchanges, state transitions between tools.
• 10 retrieval tasks (evals/federal-register-retrieval.toml): each asks one factual question about a recorded record, with a distinct answer, answerable through a recorded exchange; reward 1 if the response contains the recorded value, else 0. Plus smoke, which deploys the environment and checks replay through the MCP endpoint.
• 20 recorded episodes by openai/gpt-oss-120b:groq, zai-org/GLM-5.3-Flash:zai-org through the Hugging Face router, with their chat transcripts.
• Provenance: capture ev_e824e8d53e2ce40fe1173e27f7002aa1e3e7f80707e999d864745bac26d86575 (scoped HAR sha256 92b61ba4a50cd9beeb3af0913704c5a80536a2198a62c86f5e99d2eb56d7f5bf, 198,774 bytes), manifest view ev_23d1066c3bc138a90ac041fe3012494bc348237ced0daac373a9dc2845357f19, confidence inferred, authority historical-reference. The REA build is a prerelease of the environment commands proposed upstream in morluto/rea#4954, #4955, #4975 and #4999; the bundle installs it from the tarball it ships, so the environment runs without waiting for an npm release.

What replay means, and does not

A tool call returns the recorded response whose path and query values best match the arguments, with replay.exact_match telling you whether it matched exactly or fell back to the closest record. Reads behave like the API did on 2026-10-19. The live API keeps changing, so every answer is about the recorded snapshot, not today, and a query the session never made falls back to the closest record. The API is read-only and so is the environment: it has no state and no write tools. The manifest's unknowns say what replay cannot know.

Tasks and grading

task	tool	arguments	field	answer
qa-01-get_public_inspection_documents_by_publi	get_public_inspection_documents_by_public_inspection_documen	public_inspection_document_id=2026-18126.json	title	Airspace Designations and Reporting Points: Point Pleasant, WV
qa-02-get_documents_by_document_id	get_documents_by_document_id	document_id=2026-18019.json, fields[]=signing_date	title	Airworthiness Directives; Pratt & Whitney Division Engines
qa-03-get_suggested_searches_air_travel	get_suggested_searches_air_travel	no arguments	section	business-and-industry
qa-04-get_documents_facets_monthly	get_documents_facets_monthly	conditions[agencies][]=energy-department, conditions[publication_date][year]=2026	2026-01-01.name	January 2026
qa-05-get_documents_facets_agency	get_documents_facets_agency	conditions[type][]=PRESDOCU, conditions[publication_date][gte]=2026-09-01, conditions[publication_date][lte]=2026-09-30	executive-office-of-the-president.name	Executive Office of the President
qa-06-get_suggested_searches	get_suggested_searches	conditions[sections]=business-and-industry	business-and-industry[0].title	Automobile Safety & Fuel Economy
qa-07-get_public_inspection_documents_by_publi	get_public_inspection_documents_by_public_inspection_documen	public_inspection_document_id=2026-18126.json	toc_subject	Airspace Designations and Reporting Points:
qa-08-get_documents_by_document_id	get_documents_by_document_id	document_id=2026-18019.json, fields[]=signing_date	agencies[0].raw_name	DEPARTMENT OF TRANSPORTATION
qa-09-get_documents_facets_monthly	get_documents_facets_monthly	conditions[agencies][]=energy-department, conditions[publication_date][year]=2026	2026-05-01.name	May 2026
qa-10-get_suggested_searches	get_suggested_searches	conditions[sections]=business-and-industry	business-and-industry[0].slug	automobile-safety-fuel-economy

Episodes

model	tasks	passed	pass rate	mean tool calls	mean input tokens
openai/gpt-oss-120b:groq	10	9	90%	1.0	5,292
zai-org/GLM-5.3-Flash:zai-org	10	9	90%	1.0	6,242

Every episode is in the episodes config with its transcript; rewards are the deterministic verifier scores.

openai/gpt-oss-120b:groq missed qa-01-get_public_inspection_documents_by_publi. zai-org/GLM-5.3-Flash:zai-org missed qa-01-get_public_inspection_documents_by_publi. Misses by gpt-oss-120b:groq on qa-01-get_public_inspection_documents_by_publi, GLM-5.3-Flash:zai-org on qa-01-get_public_inspection_documents_by_publi: the model's answer did not contain the expected value.

Run it yourself

uv tool install agentenv-framework --with 'agentenv-hf @ git+https://github.com/earakely-scale/agentenv-hf-plugin@v0.3.0'
agent-env hf run earakely-scale/FederalRegisterEnv-REA --bundle federal-register --task smoke          # Docker, no model
agent-env hf run earakely-scale/FederalRegisterEnv-REA --bundle federal-register --eval federal-register-retrieval --model openai/gpt-oss-120b:groq

The retrieval eval needs a model endpoint: set [model] base_url and api_key in .agentenv/config.toml (the Hugging Face router works with any hf_ token). The bundle's agents/replay-qa is a small OpenAI-compatible tool-calling agent; any A2A agent can replace it.

How it was made

1. No instance was run: a throwaway VM called the public API directly, within the publisher's fair-access rules (a declared User-Agent, one request at a time at under 2 per second, and 69 requests in the whole session).
2. Playwright opened the API origin and ran a scripted sequence of read requests (seven agency records, 25 document reads by number covering FAA final rules, Energy, FERC and CBP proposed rules and executive orders, seven document searches, three facet counts, seven public-inspection records and three suggested searches); the HAR was scoped to the API, with headers and cookies removed.
3. rea inspect-web-network-capture retained the exchanges as Evidence; rea inspect-analysis-view projected the environment-manifest view; rea environment-export wrote this bundle; gen_tasks.py derived the retrieval tasks from the manifest and the recorded bodies; agent-env run recorded the episodes; agent-env hf publish wrote this dataset. Scripts: https://github.com/earakely-scale/rea-environments.

Compared with PortSimEnv on AgentEnv

	PortSimEnv-AgentEnv	Federal Register API (this dataset)
Environment	A simulation: a container quay, ships, cranes, weather, with A2A ship agents	A replay of a real application's API, captured from one session
Observation	Tool results computed by the simulator as the week unfolds	Recorded responses, cited to Evidence
Action space	Planning tools (check_plan, confirm_berths, advance, message_agent)	10 tools mirroring the application's endpoints
State	Full: a virtual clock, triggers, frozen windows, agents that answer	None: every call is independent; the API is read-only
Reward	Deterministic, against a CP-SAT optimum or a re-planner anchor; continuous 0 to 1	Deterministic, recorded-value match; 0 or 1 per task
What it measures	Planning under uncertainty and negotiation	Tool discovery, parameter inference and faithful retrieval in a real API
Data provenance	Real 2024 port calls and weather; simulated workloads	Real public-domain records from the Office of the Federal Register, as served on 2026-10-19
How it was built	Hand-written plugin, months of work	Generated from a capture in one scripted pass
Where it is strong	Long-horizon decisions with hard rewards	Breadth: any app you can capture becomes an environment

The two are complementary: PortSimEnv shows what a hand-built simulation can grade; this dataset shows how far a capture alone gets you, and the manifest's unknowns list exactly what a stateful version would still have to specify.

Licence and credits

Bundle code and tasks: Apache-2.0. Recorded content: responses of the public API at https://www.federalregister.gov, published by the Office of the Federal Register (U.S. public domain); the dataset repeats them as recorded. REA is MIT software by morluto; AgentEnv Framework is Apache-2.0 by Scale AI.