Maydanlabs/maydan-hf-rl-sampler

Maydan: GCC Government-Workflow Sampler for AI Agents

Maydan builds simulated GCC government-services workflows for training and evaluating AI agents. This public sampler covers Saudi GOSI-style social-insurance contribution reconciliation: 10 English-language Harbor tasks with synthetic employee and payroll records, reference solutions, and automatic checks of the final database state.

This sampler is intentionally narrower than Maydan's broader Arabic/English browser-based government-services environment work. It does not include the browser application, Absher/Qiwa/CHI residence workflows, the full Saudi country pack, held-out tasks, calibration traces, private model runs, buyer material, credentials, or production Maydan source.

What You Get

The included task family is a synthetic GOSI contribution-batch reconciliation workflow. Each task asks an agent to:

• inspect a six-employee batch
• read separate employment, contribution-history, and payroll evidence
• choose the correct contribution branch for each employee
• calculate contribution base, employer rate, employee rate, and contribution amounts
• stage all rows through a small command-line action surface
• finalize the batch
• get scored only by deterministic SQLite final-state and audit-log checks

The public action contract validates exactly these cohort values:

• new_saudi
• legacy_saudi
• non_saudi_oh_only

All names, employers, identifiers, wages, records, and database rows are synthetic. The GOSI-style records are not an official government dataset and are not legal, regulatory, or government-system advice.

Why It Exists

These tasks test stateful operational behavior: whether an agent gathers the right evidence, changes only the intended records, preserves unrelated state, and finishes with a verifiable database state.

The broader Maydan project focuses on GCC administrative workflows, including Arabic interfaces and government-service dependency graphs. This public sampler does not claim Arabic coverage, an Arabic gap, real-government fidelity, legal compliance, product readiness, or leaderboard performance. It is a small runnable compatibility sampler for the Hugging Face RL Environments direction.

Quickstart

Clone the public dataset repository into a disposable directory and run the reference/negative test suite:

mkdir -p /tmp/maydan-hf-rl-demo
cd /tmp/maydan-hf-rl-demo
git clone https://huggingface.co/datasets/Maydanlabs/maydan-hf-rl-sampler
cd maydan-hf-rl-sampler
python3 scripts/run_reference_and_tests.py

Expected final line:

ok: 10 independent F11 reference solutions and local negative tests completed

That command runs the reference solutions for all ten tasks and local negative checks for failure modes such as missing finalization, wrong branch selection, direct database writes without matching audit events, source-table tampering, and invalid cohort values. It uses local Python and SQLite only.

This quickstart was tested from a fresh public clone of commit 6a483202365ab03a6422d4535897ace93fc3e2b7. The ten seed.db files were retrieved by git clone as real SQLite databases, not as unresolved Git LFS/Xet pointer files; no separate LFS or Xet setup was needed for this test.

Harbor Compatibility

The task directories use the current Harbor task shape:

instruction.md
task.toml
environment/
solution/solve.sh
tests/test.sh

The quickstart above is a local reference/negative test command, not a Harbor agent run. Separately, the package was locally exercised with Harbor 0.24.0 and Docker/Colima using the oracle agent across all ten tasks. Each oracle run reached reward 1.0 with zero Harbor exceptions and no model/API cost. No Verifiers or OpenEnv compatibility is claimed until those frameworks round-trip this exact repository layout.

Repository Contents

README.md
LICENSE
COMPATIBILITY_REPORT.md
PACKAGE_BOUNDARY.md
RELEASE_SELECTION.md
registry.json
release_instance_manifest.json
scripts/run_reference_and_tests.py
tasks/maydan-gosi-f11-001/
tasks/maydan-gosi-f11-002/
tasks/maydan-gosi-f11-003/
tasks/maydan-gosi-f11-004/
tasks/maydan-gosi-f11-005/
tasks/maydan-gosi-f11-008/
tasks/maydan-gosi-f11-009/
tasks/maydan-gosi-f11-012/
tasks/maydan-gosi-f11-016/
tasks/maydan-gosi-f11-020/

The release includes task instructions, task config, synthetic seed databases, a tiny runtime/verifier package, reference solutions, and local negative tests. It excludes generated live.db files, logs, jobs, Python bytecode, private evidence, ledgers, credentials, and unrelated Maydan source.

Usage Signals

This repository does not install telemetry or call home. Public usage can be followed through Hugging Face repository signals exposed for dataset repositories, such as downloads, likes, and discussions. Those signals do not prove unique users, actual local executions, page views, citations, or successful task runs. Maydan's own verification and monitoring downloads may also be counted in public download totals.

If you run the sampler locally, Maydan does not receive those local execution details unless you choose to share them. For Maydan correspondence about this sampler, contact talal@maydanlabs.com.

License Scope

Copyright 2026 Maydan Labs.

Licensed under the Apache License, Version 2.0. See LICENSE.

This license applies only to this ten-task synthetic sampler boundary: the included docs, task definitions, reference solutions, tests, verifier/runtime code, and synthetic seed data. It does not license the full Maydan product, proprietary task families, held-out test sets, calibration results, buyer documents, trademarks, private data, or future country packs.

No vendored third-party libraries are included. Runtime dependencies remain external: the Dockerfiles use the python:3.13-slim base image, the sampler relies on Python standard library modules and SQLite, and Harbor is an external local runner used for validation. If Docker images or derived bundles are distributed later, preserve the applicable notices and license obligations for those dependencies and base images.