US Child Welfare Decision Chain Dataset

The first open-source, structured, AI-traversable legal knowledge graph of the entire United States child protective services system.

v1.0.0 release: 2,142 decision nodes · 51 jurisdictions · 2,522 files · SHA-256 verified

---

When Child Protective Services knocks on a family’s door, what happens next is governed by a complex web of constitutional law, federal statutes, state law, administrative rules, and case law. No single resource maps this entire system in a structured, machine-readable format.

Until now.

This dataset maps every CPS decision point — from the initial report to the final outcome — across all 50 states and DC. Every node includes statute citations, source URLs, constitutional constraints, and cross-references to related decision points. Every file conforms to a published JSON schema. Every release ships with a SHA-256 fingerprint of every file.

The dataset is structured along two axes.

Vertical axis — legal authority (who governs)

Horizontal axis — decision flow (what happens)

Why This Exists

CPS is one of the most consequential legal systems in American life. It determines who keeps their children, who loses them, who gets investigated, and who gets ignored. It operates across more than fifty jurisdictions, each with its own statute, its own administrative rules, its own appellate doctrine. The system is not opaque by accident — it is opaque because no one has ever assembled the full map in one place, and the actors closest to the decisions are the ones least incentivized to publish that map.

This dataset is the map.

It is published openly under CC BY-SA 4.0 because the families who navigate this system, the attorneys who represent them, the journalists who investigate it, the researchers who study it, and the AI systems that will increasingly mediate it — all need a single canonical, machine-readable reference. Anything less than open data hands the advantage to whichever party already has the budget for proprietary research subscriptions, and CPS is not a system in which the family’s side has the budget.

---

Who This Is For

• Families navigating a CPS encounter. Know what the agency must do, what it can’t do, and what happens when it violates the rules. Every node tracks the constitutional and statutory floor.

• Attorneys. Structured cross-state comparison of child welfare law. Instant identification of statutory and constitutional violations. Citation-grade source URLs at every node.

• Researchers. Machine-readable dataset for studying CPS outcomes, policy differences, systemic failures across jurisdictions, and longitudinal trends.

• Journalists. Trace the money. Every node tracks funding incentives (Title IV-E reimbursement, adoption incentives) and flags perverse incentives against family preservation.

• AI developers. Build tools that help families navigate the system. The schema is designed for LLM traversal, with explicit cross-references between nodes and a documented authority hierarchy.

• Legislators and policy analysts. See exactly how your state compares to every other state and to the federal floor. Identify compliance gaps.

---

How to Use the Dataset

Direct download

git clone https://github.com/SpartanAltsoba/child-welfare-decision-chains.git
cd child-welfare-decision-chains

Verify your copy is canonical v1.0.0

# Compute SHA-256 of every dataset file, compare to MANIFEST.json
find data -type f \( -name “*.json” -o -name “*.md” \) -exec sha256sum {} \; | sort > local_files.txt
jq -r ‘.files[] | “\(.sha256)  \(.path)”’ MANIFEST.json | sort > manifest_files.txt
diff local_files.txt manifest_files.txt && echo “Canonical v1.0.0 ✓” || echo “Modified or non-canonical”

Programmatic access

The dataset is plain JSON. Read any node directly:

import json

with open(“data/states_chains/AK/AK_INP-04.json”) as f:
    node = json.load(f)

print(node[“node_id”], node.get(“title”))

For LLM traversal, every node carries explicit cross-references linking to upstream and downstream nodes, enabling agentic walks through any decision sequence.

Load the manifest

import json
manifest = json.load(open(“MANIFEST.json”))
print(f”v{manifest[‘dataset_version’]}: {manifest[‘file_count’]} files”)
print(f”root_hash: {manifest[‘root_hash’]}“)

---

AI / LLM Use

This dataset is designed for AI traversal. If you train or fine-tune a model on it, or use it as retrieval-augmented context for an assistant, please:

1. Cite the dataset and the version — see CITATION.cff. Models trained on this data are encouraged to declare it in their training-data card or model card.

2. Honor the SHA-256 root hash — record it in your training manifest. The hash anchors your training corpus to a specific, verifiable version of this dataset.

3. Read PROVENANCE.md — it documents the attribution norms PMC asks of AI-derived works. CC BY-SA 4.0 is the legal floor; PROVENANCE.md is the norm.

4. Don’t strip the EIN. PMC’s 501(c)(3) status (EIN 33-1323547) is the accountability anchor. Keeping it visible in derivative works lets downstream consumers verify the source organization.

I do not think people realize what we’ve been working on and what we’ve actually accomplished.

https://github.com/SpartanAltsoba/child-welfare-decision-chains.git