Better safety guidance without centralizing private case data.

It hides in job posts, in chat threads on a recruiter’s phone, in fee receipts, in a contract clause that contradicts a labour ministry circular published last week. The signal is real but it is fragmented across posts, chats, documents, fees, threats, and shifting laws across many corridors.

The institutions that could act — platforms, NGOs, regulators, embassies, courts — each see a slice. Workers, who carry the most context, often have the least bandwidth, the slowest connection, and the highest cost of being wrong.

NGOs face the same fragmentation. One organization may see a fee pattern in shelter intake, another may see the same recruiter language in a hotline transcript, while a regulator only needs the anonymized corridor-level trend. Today those signals are hard to share safely, quickly, and with enough provenance for another institution to trust them.

The pattern is clear: useful safety knowledge already exists in public sources, casework experience, and academic research. What’s missing is the connective tissue that makes that knowledge available at the moment a worker, caseworker, or moderator needs it, without funnelling private cases into one centralized warehouse.

A useful rule of thumb for where AI gets dramatically better, fast:

capability spike ≈ verifiability × training attention × data coverage × economic value

Migrant worker exploitation scores low on every factor. Outcomes are hard to verify (workers go silent, cases close years later, and definitions vary by jurisdiction). Training corpora rarely emphasize labour law, recruitment fees, or corridor-specific scams. Public data is thin and scattered across regulators, NGOs, and court filings. The economic incentive to optimize a frontier model for this domain is small relative to coding or search.

The result: stock models give plausible-sounding answers that are wrong in load-bearing ways. They miss the cited statute. They under-estimate corridor risk. They invent NGO contacts. They paraphrase a recruiter euphemism instead of flagging it.

Until inherent capability arrives, the gap has to be closed by structure: deterministic grep rules that fire before the model speaks, retrieval against vetted corridor packs, tool calls that ground claims in verified sources, and a harness ecosystem that can refuse, narrow, anonymize, verify, evaluate, or train around an answer when needed. Gemma 4 was chosen because it ships with native tool-calling and can run in local, edge, and on-device-oriented deployments, which is what makes the harnesses practical.

DueCare publishes vetted knowledge packs, a tools registry, evaluation suites, and a small AI harness ecosystem. Anyone can run them. Everyone can verify them. Citation-critical paths are designed to fail closed or narrow the answer when they can’t cite a public source.

The mission isn’t a single product. It’s shared infrastructure: a thin public hub that holds reviewed knowledge, paired with private deployments at the edges where worker context lives.

That infrastructure also gives NGOs a safer way to coordinate with each other and with regulators: share vetted public sources, pattern labels, contact updates, and anonymized corridor trends, not raw case files, names, documents, or narratives.

DueCare is not affiliated with the institutions below. It is built so that NGOs, embassies, labour ministries, and academic centres can use vetted packs and the harness to amplify what they already do, not to replace it.