Skip to content

Credits and attributions

DueCare is built on a stack of open-source software, public legal sources, and prior research. This file enumerates the dependencies, data sources, and influences a judge or contributor should be aware of, with attribution.

Model and inference

  • Gemma 4 — Google DeepMind. Open-weight base model used as the local runtime across all three Kaggle kernels and the sibling Android app. Licensed under Apache 2.0; the Gemma Prohibited Use Policy also applies. We use the E2B, E4B, 26B-A4B, and 31B-IT variants from google/gemma-4 on Kaggle.
  • Unsloth — unsloth.ai / github.com/unslothai/unsloth. Used for LoRA fine-tuning of Gemma 4 in the A-00 control plane (DueCare Fine-tuning and Evaluation). Apache-2.0.
  • MediaPipe LiteRT — Google AI Edge. Powers Gemma 4 E2B inference fully on-device in the sibling Android app (duecare-journey-android v0.9.0). Apache-2.0.
  • llama.cpp — github.com/ggml-org/llama.cpp. Reference deployment target for the GGUF export path. MIT.

Runtime libraries

  • FastAPI — the live-demo server is a FastAPI app (packages/duecare-llm-server). MIT.
  • Pydantic v2 — data-model layer across every harness contract. MIT.
  • Uvicorn — ASGI server for the kernel-embedded live demo. BSD.
  • DuckDB — local storage for the live demo. MIT.
  • Cloudflare quick tunnel — every Kaggle kernel prints a public https://*.trycloudflare.com URL for the reviewer-facing demo.
  • Hugging Face Transformers + PEFT + TRL — used through Unsloth's fine-tune path. Apache-2.0.

Kaggle community

  • @bwandowando — kaggle.com/bwandowando. The install-marker-then-restart sequencing that lets a heavy, C-extension-bearing stack (torch + triton + Unsloth) be installed inside a running Kaggle session comes from his published recipe. We follow the ordering faithfully in kaggle/02-live-demo/kernel.py and kaggle/_archive/notebooks/A-07-bench-and-tune/kernel.py, because the ordering is the part that actually works. Without it the Gemma 4 kernels would not boot on Kaggle.
  • Daniel Han and the Unsloth team — the pinned Gemma 4 + Unsloth version set installed by those kernels is theirs.

Models evaluated in the benchmark

The DueCare Harness-Lift Benchmark measures paired lift — the same prompt, with and without the harness, graded identically. That design means other organisations' models are run as measurement subjects, and several are also used as judges. All of them deserve naming. Nothing here is a claim about a vendor's product quality in general; the metric is how much a prompt-level harness changes a response on one migrant-worker safety rubric.

Evaluated as subjects (7 on the public board):

  • Google DeepMind — Gemma 4 (gemma4:31b plus the E2B / E4B / 26B-A4B variants). Apache License 2.0.
  • OpenAI — gpt-oss:20b, gpt-oss:120b (open-weight release), and gpt-4o / gpt-4o-mini in the wider frontier study.
  • Zhipu AI / Z.ai — glm-5.1, glm-5.2.
  • DeepSeek — deepseek-v4-pro and earlier v3.x releases.
  • Alibaba Qwen — qwen3.5:397b, qwen3-coder:480b, qwen3-next:80b.
  • MiniMax — minimax-m2.7.
  • Moonshot AI — Kimi K2 / K3 lanes.
  • Anthropic — Claude models in the frontier comparison lanes.
  • Mistral AI — mistral-large in earlier generation runs.

Used as judges — deepseek-v4-pro, glm-5.2, gpt-oss:120b: a three-model panel with self-family exclusion, so no model grades its own family. Judge disagreement is reported rather than hidden.

Serving and routing

  • Ollama — ollama.com / github.com/ollama/ollama. Local and cloud serving for the open-weight models in the benchmark engine. MIT.
  • OpenRouter — openrouter.ai. Routing layer for closed frontier models in the comparison lanes.
  • Hugging Face Hub — model distribution and the huggingface_hub client. Apache-2.0.

Typography

The workbench, hub, and slide surfaces load these via Google Fonts. All three families are SIL Open Font License 1.1:

Research tooling with citation requests

Some libraries are permissively licensed but their authors ask to be cited academically. Honouring that:

  • VADER (vaderSentiment) — Hutto, C.J. & Gilbert, E.E. (2014). VADER: A Parsimonious Rule-based Model for Sentiment Analysis of Social Media Text. Eighth International Conference on Weblogs and Social Media (ICWSM-14). Used for tone analysis in the prompt/response NLP notebooks. MIT.
  • textstat — readability metrics in the prompt/response NLP notebook. MIT.
  • sentence-transformers — Reimers, N. & Gurevych, I. (2019). Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks. EMNLP
  • Apache-2.0. The specific checkpoints used are listed in LICENSES.md.

These are public sources. The knowledge-pack RAG corpus quotes section numbers, captions, and citation strings verbatim with full attribution to the issuing authority. The list below is illustrative, not exhaustive — the full inventory lives in packages/duecare-llm-chat/src/duecare/chat/harness/_citations.json.

  • International Labour Organization (ILO) — Conventions C029 (Forced Labour), C181 (Private Employment Agencies), C189 (Domestic Workers); Forced Labour Indicators; Global Estimates of Modern Slavery 2022; Profits and Poverty: The Economics of Forced Labour 2024.
  • United Nations — Palermo Protocol (Protocol to Prevent, Suppress and Punish Trafficking in Persons).
  • Philippines — RA 8042 (Migrant Workers and Overseas Filipinos Act); RA 10022 (amendment); RA 10361 (Batas Kasambahay); RA 9208 (Anti-Trafficking); RA 11862 (recent amendment); POEA Memorandum Circular 14-2017 / 02-2007 (now DMW policy); DMW model contract.
  • Indonesia — BP2MI Regulation 9/2020.
  • Nepal — Foreign Employment Act 2007 §11(2); 2015 Free-Visa- Free-Ticket Cabinet Decision.
  • Bangladesh — Overseas Employment Act 2013.
  • Hong Kong — Employment Ordinance Cap. 57; Money Lenders Ordinance Cap. 163; Employment Agency Regulations Cap. 57A.
  • Singapore — Employment of Foreign Manpower Act Cap. 91A §22A.
  • United Arab Emirates — MoHRE Decree 765/2015.
  • California (US) — Civil Code §1714(a), the general duty-of-care doctrine that gave the project its name.

Influences and prior art

  • Taylor Amarel's 2025 Kaggle red-teaming research — LLM complicity in modern slavery: from native-blind... (Kaggle write-up). The empirical evidence that frontier LLMs consistently fail on migrant-worker safety prompts. Motivates the harness design.
  • Sister synthetic-evidence harness — a fully synthetic test-data generator (synthetic_test_evidence/) with closed-set indicator vocabulary, prompt packs (v1 lenient, v2 chain-of-thought + strict thresholds), reproducible release zip, deterministic CI smoke runner, and per-lane sample cases. Maintained alongside DueCare; outputs are watermarked and safe for any public-cloud test environment.
  • DueCare Journey for Android (sibling app) — github.com/TaylorAmarelTech/duecare-journey-android v0.9.0. On-device worker-facing app demonstrating the LiteRT path for Special Technology Track eligibility.
  • Partner NGOs — Polaris Project, International Justice Mission (IJM), ECPAT, the Philippine Overseas Employment Administration (POEA, now DMW), BP2MI (Indonesia), HRD Nepal. Their published guidance, statutes, contact pathways, and case archetypes shaped the GREP rule library and RAG knowledge packs.
  • Cal. Civ. Code §1714(a) doctrine — March 2026 California jury verdict against Meta and Google for defective platform design. We apply the same duty-of-care standard to language models.

Directory data (organisations that help workers)

configs/duecare/research_monitor/migrant_support_orgs.yaml compiles contact routes for migrant-support organisations — helplines, shelters, legal aid, unions, resource centres, labour attachés, and intergovernmental bodies — so the harness can point a worker or caseworker at real help instead of inventing a number.

  • Every entry is published organisational contact information. No individual's personal details are included, and the catalog header states that constraint.
  • Per-entry provenance is recorded in each row's notes field, naming the official directory or publication the contact came from (for example a ministry labour-wing directory or an agency's own published hotline page).
  • The organisations are credited by name in the data itself and are the authors of their own contact information; DueCare claims no ownership over it and asserts no endorsement by them.
  • url_verified marks whether a link has been re-checked. Contact details go stale, which is exactly why they live in a versioned knowledge object rather than being trained into model weights.

If you represent an organisation listed here and want an entry corrected or removed, open an issue on the repository and it will be changed.

AI-assisted development

Stated plainly because the repository makes it visible anyway, and because a project whose thesis is "real, not faked" should not be coy about its own tooling.

DueCare was built with substantial AI coding assistance — principally Anthropic's Claude (via Claude Code) and OpenAI's Codex, whose working notes, review prompts, and handoffs are committed under docs/ and .claude/ rather than scrubbed. Architecture, research direction, domain expertise, the benchmark design, and every published claim are the author's. AI assistance does not extend to the results: benchmark numbers come from recorded runs that regenerate from (git_sha, dataset_version), and where a result did not support a claim it is published as a negative result instead.

Synthetic content disclaimers

Every shipped sample bundle (packages/duecare-llm-chat/src/duecare/chat/static/samples/*) is fully synthetic and carries a SYNTHETIC TEST DATA - NOT REAL watermark or banner. Composite agency names ("Sunburst Manpower Services", "HK Domestic Jobs", "MetroMed Diagnostic Center") are labelled (composite) in any judge-visible surface. No real worker PII, no real recruiter PII, no real case data is included in this repository or any of its samples.

License

The DueCare project code is MIT-licensed. The Gemma 4 model weights are licensed under Apache 2.0, with the Gemma Prohibited Use Policy also applying. Library dependencies retain their respective licenses (see each library's repository). Knowledge-pack content quotes from public statutes and circulars is used under fair-use citation; redistribution of those quoted fragments inside the knowledge packs is permissible because the underlying texts are public-record government and IGO publications.

How to add a citation

If a knowledge pack, GREP rule, or response template adds a new statutory or organizational citation, add it to:

  1. The relevant pack JSON (packages/duecare-llm-chat/src/duecare/chat/harness/_citations.json or the equivalent for the specific harness).
  2. The "Legal and policy sources" section of this file.
  3. The relevant test fixture if a contract test pins it (e.g. packages/duecare-llm-server/tests/test_slides_surface.py asserts RA 8042 is the migrant-workers statute, not RA 11227).