Skip to content

Copy-ready final Kaggle post: DueCare closeout and node-first future

This is the final project-update post to publish in the Kaggle community as Gemma 4 Good judging concludes. The text below is intentionally self-contained, dated, and candid about completed work, planned provider evaluation, missing human evidence, and the post-competition hosting change.

Kaggle's supported CLI can browse/read forums and manage datasets, models, competitions, and notebooks; its documented forum commands do not publish a community discussion. Copy the title, body, and tags below into Kaggle's discussion editor without changing the evidence labels.

Title

DueCare after Gemma 4 Good: a reusable capability-gap architecture, a node-first future, and an honest final status

Post body

Hi Kaggle community,

As Gemma 4 Good judging winds down, I wanted to leave one final DueCare update: what was built, what changed after submission, what still has not been proven, and how the project will remain available after its centralized demo hosting is retired.

DueCare began as a Gemma 4 safety project for migrant-worker protection. The core problem was that a general model could sound reasonable while missing the specific indicator, rule, evidence path, or safe next step that a worker, caseworker, platform reviewer, or regulator actually needed. The project grew into a modular system around the model: deterministic pattern checks, versioned retrieval, tools, citations, privacy controls, traces, evaluation, notebooks, review packets, and public knowledge-sharing boundaries.

The most important architectural lesson now has its own public blueprint:

https://tayloramareltech.github.io/duecare/architecture/capability_gap_blueprint/

It describes a pattern that can be rebuilt for another industry or another LLM capability gap. The reusable parts are the separation between seven planes: claims, evidence, runtime harness, evaluation, human governance, public network/publication, and agent operations. The trafficking-specific rules are not assumed to transfer. A new domain still needs its own authoritative sources, rubric, benign controls, privacy policy, escalation paths, and qualified reviewers.

What was added or tightened after submission

  • The public project-status page, maintainer handoff, Claude Code handoff, transition plan, repository maps, and GitHub Pages documentation now agree on one current state.
  • A backend-free continuity build exports all 51 public website routes and five checksum-bound public snapshots from the real FastAPI templates. Mutable APIs, accounts, submissions, automation, and private state fail closed.
  • Website layout defects were repaired and the current status page was checked in real desktop and emulated-mobile rendering, including the new architecture and evaluation cards.
  • The provider budget is now enforced before seven registered model transports: four primary generation routes, adverse-media verification, model-failure candidate generation, and contextual judging. Attempts, token ceilings, and reviewed worst-case cash are reserved atomically; retries are new attempts.
  • The evaluation runner now supports a publication-grade one-dimension-per-call protocol and an economical one-call holistic protocol, real DueCare GREP/RAG/tool context, response/context/rubric hashes, resumable results, and explicit cross-family versus self-family labels.
  • Hermes, server automation (including the legacy OpenClaw-compatible name), the Public Information Research Monitor, outreach drafting, and human review are documented as different things. Agents may propose and route work; they cannot promote their own output into truth or claim that civil-society review happened.

Kimi K3 and contextual judging

I also froze a 500-item directional Kimi K3 study instead of leaving “test a frontier model” as a vague next step.

The exact topology is:

  1. 500 baseline Kimi K3 candidate answers;
  2. a local deterministic DueCare grade for every successful answer;
  3. 500 Gemini 3.1 Pro cross-family judgments with the same versioned DueCare context; and
  4. 500 separately labeled Kimi K3 contextual self-judgments.

That is a maximum of 1,500 hosted calls. The frozen worst-case reservation is 7,296,582 estimated input tokens, 1,152,000 maximum output tokens, and US$34.448916 at the rates checked on 2026-07-28.

This campaign has not run. Five earlier bounded Kimi baseline-access requests and a later two-arm baseline/full-harness smoke reached Ollama but returned HTTP 402 because the account had no extra-usage balance, and no Gemini API credential is present. Therefore there are zero Kimi candidate completions, zero Gemini judgments, zero Kimi self-judgments, and zero new human ratings. Access failures are not benchmark results. The paired smoke did fix a real full-harness tool-call adapter bug and left an exact tested resume receipt; Kimi execution is now a funded-later next step rather than a vague or blocking closeout item. The run instructions are public here:

https://tayloramareltech.github.io/duecare/research/model_failure_run_readiness/

Paired smoke receipt:

https://github.com/TaylorAmarelTech/duecare/blob/master/configs/duecare/benchmarks/kimi_k3_harness_lift_smoke_20260728.json

The provider references used to freeze that plan are the official Kimi K3 cloud catalog, Gemini 3.1 Pro Preview model documentation, and Gemini API pricing. Pricing and access can change, so a future run must recheck all three before unlocking its ledger:

Kimi's self-judgment would be a bias diagnostic, not independent validation. Gemini is the primary cross-family automated judge for this directional pilot. Neither is a substitute for qualified human review.

What remains incomplete

The human-validation packet contains 364 review items across 182 strata, but it still has zero qualified independent ratings. The training-data diversification workbook defines 75 exact risk, benign-neighbor, and counterfactual slots, but zero content rows have been admitted because source-rights snapshots and two-person adjudication are still missing. Those gaps remain visible rather than being filled with synthetic approvals.

High-value next work, if a future maintainer or partner reopens it, is:

  • fund and run the frozen Kimi/Gemini phases under their separate hard budgets;
  • test Meta Muse Spark 1.1 as another explicitly versioned comparison lane;
  • send the disagreement slice, not every easy row, to qualified reviewers;
  • complete multilingual, code-switch, benign-control, corridor-diversity, and temporal legal-freshness review; and
  • operate real partner nodes with consented governance instead of treating a public outreach form as an active civil-society network.

From a central demo hub to durable pages and deployable nodes

Render remains available during competition grading. After grading is confirmed complete, the plan is to retire the centralized Fly/Render-style demo service and preserve the public presentation on GitHub Pages:

That change means the main public site will no longer function as a centralized mutable hub. Public submissions, accounts, admin state, automation, and live hub APIs will be unavailable on the static site. This is intentional, not an outage disguised as continuity.

The software itself remains deployable. An NGO, regulator, researcher, platform team, or developer can run a local or organization-owned DueCare node from the repository's Docker/local deployment paths, keep sensitive material inside that trust boundary, and exchange only reviewed packs, sanitized fact objects, aggregate signals, or signed/hash-bound artifacts. A partner that needs mutable coordination can self-host the FastAPI hub rather than depending on one demonstration server.

The post-competition model is therefore less centralized and more consistent with the privacy premise: durable public knowledge and reproducibility on GitHub/Kaggle, with operational intelligence handled by independently governed nodes.

Start here

Thank you to everyone who reviewed, tested, challenged, or learned alongside the project. If someone picks DueCare up later, the goal is that they inherit not only code, but the decisions, boundaries, reproducible plans, and honest unknowns needed to improve it safely.

  • Taylor

Suggested tags

gemma-4-good responsible-ai llm-evaluation rag ai-safety human-trafficking privacy open-source local-ai kaggle-notebooks

Use these existing public, synthetic/reviewer-safe assets in this order:

  1. kaggle/01-duecare-exploration-workbench/tests/test-results/screenshots/desktop-chromium-harness-tiles.png
  2. caption: “DueCare's independently toggleable Persona, GREP, RAG, Tools, Online, and Import capabilities in the Kaggle workbench.”
  3. kaggle/01-duecare-exploration-workbench/tests/test-results/screenshots/desktop-chromium-model-picker.png
  4. caption: “The shared workbench model service and six public user lanes.”
  5. packages/duecare-llm-chat/src/duecare/chat/static/synthetic/receipt_PH_HK_001.png
  6. caption: “Synthetic fee-camouflage evidence used to test substance-over-form recognition; no real worker data.”

Publication check

Immediately before pasting:

  1. Confirm judging has concluded or change only the opening phrase to “as judging continues”; do not imply an official date that Kaggle has not posted.
  2. Confirm the three Kaggle notebook URLs and both GitHub Pages URLs return 200.
  3. Keep the Kimi/Gemini paragraph at zero results unless the frozen campaign has actually executed and its report passed claim/privacy review.
  4. Keep Render active until the separately documented transition gate passes.