DueCare notebooks, datasets, and the reusable kit¶
A single catalog of the public DueCare surfaces on Kaggle: the analysis and
applied notebooks, the open datasets they read, and the importable
duecare-llm-kit package. Every analysis notebook is CPU-safe and recomputes
its figures live from a public dataset -- no hidden state, attach the dataset
and re-run any cell.
How to read the status column¶
- live -- published and reachable. It is linked below and from the public hub at duecare-ai.com/data and duecare-ai.com/kernels.
- queued -- a build script or owner-side draft exists, but there is no verified public URL. Queued is not a publication promise. The item gains a link only after the public page returns successfully and the run/artifacts meet the claim made here.
Slugs follow https://www.kaggle.com/code/taylorsamarel/<slug> for notebooks
and https://www.kaggle.com/datasets/taylorsamarel/<slug> for datasets.
Live reconciliation on 2026-07-27 used the authenticated Kaggle command line
plus unauthenticated public-page checks. The owner listing contains many legacy
and private/unpublished drafts, so owner listing alone is not evidence that a
notebook is public. Direct checks confirmed the Prompt Intent and Deterministic
Verification notebooks are public and COMPLETE; the active/optional kernel
status boundary is maintained in
current_kaggle_notebook_state.md.
The same closeout pass validated all three active script-kernel sources, all five active/optional root kernels, generated page-source references, and 31 focused Kaggle/notebook contract tests. Those are offline source and execution- contract checks; they do not relabel canceled Kaggle runs as complete and did not dispatch a new remote run.
Analysis¶
The harness-lift result, opened up. These read the public grade and text datasets and recompute every figure.
| Status | Notebook | What it is |
|---|---|---|
| live | Start Here: Harness-Lift Benchmark | The benchmark front door: the headline lift and the cross-model board, recomputed live, with a guided tour of the whole collection. |
| live | Does A Safety Harness Help? (flagship) | The publication-grade walk-through: thirteen sections, dozens of live charts, the +40.7 headline, and the honest counter-evidence. |
| live | Per-Dimension Grades Explorer | The exhaustive one-judge-call-per-dimension sweep: per-dimension A-E lift, sliceable by model and judge. |
| live | Cross-Model Leaderboard Deep-Dive | Every model ranked by raw lift and by ceiling-adjusted normalized gain, so a strong baseline is compared fairly with a weak one. |
| live | Prompt And Response NLP Explorer | Text analytics over the prompt/response showcase: length, distinctive vocabulary (baseline vs harnessed), refusal and citation markers, readability. |
| live | CoT Reasoning Explorer | Browse the chain-of-thought reasoning traces the harness produces, prompt by prompt. |
| live | CoT Reasoning Analysis | Quantitative analysis of the reasoning chains: structure, length, and indicator / citation density. |
| live | Harness Grades Data Card | Schema, provenance, and coverage of the grades panel, and how to load it. Read this before trusting the charts. |
| live | CoT Reasoning Data Card | Schema and provenance of the chain-of-thought dataset. |
| live | Prompt Intent And Attack Explorer | The attack taxonomy: intent, framing, and category coverage of the adversarial prompt set. |
| live | CoT Direction And Intent Explorer | Where each reasoning chain points: direction, intent, and refusal geometry. |
| live | Corridor And Sector Atlas | Lift mapped across migration corridors and labor sectors, so you can see it holds beyond one geography. |
Benchmark deep-dives reached from the Start Here index (per-claim analyses over the same grades panel): reproduce the harness lift, where the harness helps most, statistical robustness, plus judge-agreement, methodology-and-controls, benchmark-convergence, impact-and-coverage, and benchmark-as-training-signal (all reachable from the Start Here index above).
Applied use cases¶
Offline, paste-in workflows. No dataset required; they run as pure local Python around the harness primitives.
| Status | Notebook | What it is |
|---|---|---|
| live | NGO Case Triage | Paste a worker account, get an ILO-grounded triage: indicators, risk level, evidence gaps, next steps, referrals, and a draft note. |
| live | Worker Self-Check | A worker pastes a suspicious message and gets a plain-language warning and next steps. |
| live | Platform Moderation At Scale | Screen risky recruitment posts and ads into a review queue with a reason for every decision. |
| queued | Chain Of Thought Generator (duecare-chain-of-thought-generator) |
Turn a prompt into a structured, ILO-grounded reasoning chain. |
| queued | Prosecutor Case Building (duecare-prosecutor-case-building) |
Draft case-building workflow; present in the owner listing but its public URL returned 404 on 2026-07-27. |
| queued | Regulator Compliance (duecare-regulator-compliance) |
Compliance-monitoring view for labor ministries and regulators: corridor rules, fee caps, and an evidence trail. |
| queued | Developer Integration (duecare-developer-integration) |
The software-to-software path: call the harness from your own code, structured request in and structured analysis out. |
Advanced¶
| Status | Notebook | What it is |
|---|---|---|
| queued | The Entire System (duecare-the-entire-system) |
End-to-end tour of the whole DueCare substrate: runtime, harness layers, knowledge, training, and judging. |
| queued | Semantic Landscape (duecare-semantic-landscape) |
An embedding-space map of the prompt and knowledge corpus: clusters, gaps, and coverage. |
| queued | Cross-Industry Capabilities (duecare-cross-industry-capabilities) |
The same harness across domains beyond trafficking (tax evasion, financial crime, and more). |
Knowledge¶
| Status | Notebook | What it is |
|---|---|---|
| queued | Knowledge Base Explorer (duecare-knowledge-base-explorer) |
Browse the GREP rules, the RAG corpus, the ILO instruments, and the corridor fee-caps behind the harness. |
| queued | Getting Started (duecare-getting-started) |
A one-page hub: set DueCare up in minutes, plus a catalog of every published surface. |
| queued | Fact Check And Reproducibility (duecare-fact-check-and-reproducibility) |
Verify the headline numbers and reproduce them from the public data. |
| live | Deterministic Verification | Offline verification tutorial; anonymous page returned 200 and authenticated status was COMPLETE on 2026-07-27. |
Publication queue and stopping decision¶
No additional notebook needs to be published merely to make the repository look complete. The public catalog already covers the reviewer path, benchmark analysis, applied examples, training-data inspection, and the optional Kaggle Community Benchmark. Bulk-publishing private drafts would add maintenance and link debt without adding evidence.
The queued notebooks remain useful candidates. Advance one only when it closes a named audience or evidence gap, then:
- regenerate it from its checked-in builder and run its deterministic/static assertions;
- verify every attached dataset version, license, privacy boundary, and claim;
- push one notebook deliberately, wait for
COMPLETE, inspect the saved outputs, and confirm the public page returns HTTP 200 without credentials; - add its link here and on the website in the same change; and
- keep it auxiliary unless root
AGENTS.mdandkaggle/_INDEX.mdexplicitly change the three-surface active reviewer path.
Datasets¶
Open-licensed. Grades and scores carry no response text and no personal data; the text datasets are synthetic composite scenarios with kernel metadata scrubbed and a conservative PII scan applied.
| Status | Dataset | What it is |
|---|---|---|
| live | Harness Benchmark Grades | The judged panel: one 0-100 score per (model, arm, prompt, judge) plus five A-E components. 85,417 grade rows across 7,973 prompts and 8 models. The primary evidence file. |
| live | Harness Per-Dimension Grades | Higher resolution: one reasoned judge call per rubric dimension. Re-versioned as the exhaustive sweep grows. |
| live | Cross-Model Harness Leaderboard | A citable flat CSV of the cross-model board: baseline, harnessed mean, raw lift, normalized gain, and win rate per model. |
| live | Harness Lift Controls | The placebo, negative-control, and applicability-audit results -- including the honest, inconclusive parts. |
| live | Prompt Response Showcase | The raw adversarial prompt plus the model's baseline, harness-core, and harness-full answers side by side. 1,087 prompts x 3 responses. |
| live | CoT Reasoning | Multi-perspective chain-of-thought reasoning traces grounded in the ILO forced-labour indicators. 2,020 chains, 1,740 train / 280 held out. |
Training corpora (measured-response, multiperspective, the proof corpus, and the adapter / byte-model learning studies) are catalogued on the hub's Data & downloads page.
Reproduce the headline in a few lines¶
Attach the grades dataset to a Kaggle notebook (or download the CSV) and read it with pandas:
import pandas as pd, glob
grades = pd.read_csv(glob.glob("/kaggle/input/**/panel_grades.csv", recursive=True)[0])
piv = grades.groupby(["prompt_id", "arm"])["score_0_100"].mean().unstack() # mean the judge panel
lift = (piv["harness_core"] - piv["baseline"]).dropna() # pair harnessed vs baseline
print(f"mean lift +{lift.mean():.1f} over {len(lift):,} paired prompts, {100 * (lift > 0).mean():.1f}% improved")
These are benchmark response-quality results under an LLM judge panel, not field-detection metrics. Source and build scripts: github.com/TaylorAmarelTech/duecare.
The reusable kit: duecare-llm-kit¶
The indicator engine, chart helpers, HTML-report generator, corpus exporter, and deterministic verifier that ship embedded in these notebooks -- packaged as importable Python so you can use them without opening a notebook.
No DueCare distribution is published on PyPI yet. From a repository checkout:
from duecare.kit import scan, generate_report
for hit in scan("The agency took my passport and I have not been paid."):
print(hit["indicator"], "--", hit["ilo_ref"])
generate_report("panel.jsonl", "duecare_report.html") # self-contained HTML lift report
Full API, CLI, and usage snippets: packages/duecare-llm-kit/README.md.