Skip to content

DueCare Roadmap

Current as of 2026-07-28.

This is the strategic roadmap. The dated closeout receipt is authoritative for the 11 inherited decisions. The generated DEFERRED_WORK.md register contains zero current items and becomes authoritative only for specifically reopened work.

Historical and topic-specific plans remain as provenance under docs/research/ and docs/_archive/; they do not authorize model calls, publication, or the promotion of candidate data.

Current Stopping Point

  • The source, Render production website, independent read-only continuity Pages copy, MkDocs Pages documentation, active notebook sources, package build contract, and deterministic verification stack are maintained from reviewed master revisions with distinct deployment ownership.
  • Render remains production through competition grading. After grading is owner-confirmed complete, the approved event-gated target is durable Pages presentation plus independently governed runtime nodes; Pages will not preserve mutable hub APIs. Follow POST_COMPETITION_HOSTING_TRANSITION.md.
  • The active Kaggle submission sources are exactly 01, 02, and A-00. The 03 and 04 benchmark surfaces are optional and are not prerequisites for repository closure.
  • All 18 Python distributions build and clean-install under the release contract. No DueCare distribution is claimed as published to PyPI; first publication was explicitly declined for closeout.
  • The model/flywheel stack is deliberately cost-stopped. Four sentinels are present, five recurring Windows tasks are disabled, and a finite provider budget is required before any caller resumes.
  • The exhaustive generation phase is complete. Exhaustive per-dimension closure was declined: 47,813 of 708,471 panel cells are present and 660,658 are still missing in the dated receipt. It remains partial experimental evidence, not part of the default comparable board.
  • The training scope remains deliberately red and excluded from closeout claims. Reopening it would require human curation/adjudication of the 75-row workbook, a clean quality audit, and refreshed append-only provenance.

The model-free release command is:

$env:DUECARE_MAX_PLANNED_MODEL_CALLS = '0'
python scripts/validate_publication_readiness.py --scope core

Near-Term Maintenance

The registered legacy Ruff slice completed on 2026-07-28: the three selected files pass the configured rules without suppressions, backed by their offline behavior tests. No ready_local item remains in the canonical queue. Seven registered transports now have shared atomic budget coverage: four primary generation paths, adverse-media verification, model-failure candidate generation, and contextual judging. Other callers remain future bounded migrations under the exact coverage statement in PROVIDER_BUDGETING.md.

Conditional Evidence And Dataset Opportunities

If a receipt reopen condition is met, new evidence should deepen validity rather than inflate surface count:

  1. Complete the source-bound 75-row corridor-diversification workbook with two independent curators and native-language review where required.
  2. Create a qualified human gold set and measure judge-human agreement before using model-judge scores as a stronger validity claim.
  3. After an explicit budget approval, run a small frozen smoke matrix with immutable model IDs, prompts, rubric, harness, decoding, and cash/token caps. Kimi K3 and Meta Muse Spark 1.1 are required comparison lanes; record an inaccessible provider lane rather than silently substituting another model. A five-attempt Kimi K3 access check on 2026-07-28 returned HTTP 402 with zero completions or provider-token usage, so Kimi remains unavailable until the billing owner deliberately funds extra usage; the 500-prompt run did not start. A separate two-attempt baseline/full-harness smoke reached the same access stop and produced no score, while fixing and testing the full-harness tool adapter. Resume the exact hash-bound pair in kimi_k3_harness_lift_smoke_20260728.json before any larger Kimi run.
  4. Continue the isolated per-dimension lane only from its resumable coverage receipt; never merge its incomplete metrics into the default board.
  5. Version datasets append-only. Preserve source rights, checksums, lineage families, quarantine outcomes, and split-isolation evidence with each release.

No extra Kaggle notebook is needed merely to make the project look complete. Publish or rerun a notebook only when it carries a distinct, reviewable evidence artifact that an existing active surface cannot express.

Conditional Product And Integration Opportunities

If a future maintainer deliberately reopens product work, the strongest extensions are:

  • a stable domain-pack and harness-plugin contract with one minimal reference implementation;
  • measured on-device behavior on a frozen Gemma revision, including latency, memory, quantization, and safety deltas;
  • a worker-facing multimodal review path that keeps raw documents local and exposes provenance and trust boundaries;
  • bounded integration recipes for NGO, regulator, platform, and research deployments; and
  • a versioned, curator-approved knowledge-refresh workflow that stages changes but never publishes autonomously.

These are future product programs, not claims about the current release.

Community And Research Opportunities

  • Calibrate the benchmark with anti-trafficking and migrant-rights specialists.
  • Publish a citable dataset or package only after the owner selects a release disposition, license, support boundary, and exact tagged artifact.
  • Use the optional Kaggle Community Benchmark surface only if the account owner wants to host a maintained benchmark and can commit to moderation and version governance.
  • Explore cross-domain ports as separate, source-gated evidence lanes. Do not reinterpret a synthetic seed as jurisdictional coverage.
  • Retain negative results and no-lift findings; they are part of the scientific record, not cleanup candidates.

Long-Term Maintenance Rules

  • Stable behavior belongs in code, tests, and versioned knowledge objects. Volatile rules, contacts, fees, office names, and advisories require dated source review rather than memorization in training targets.
  • Entity intelligence remains propose-only and curator-reviewed, separate from worker-facing decisions and the GREP/RAG knowledge layer.
  • Every autonomous or scheduled path stays fail-closed on privacy, budget, provenance, and publication authority.
  • Claims stay attached to exact artifacts and dates. A passing core gate does not make the intentionally separate training lane green.

How To Update This Roadmap

Preserve the dated receipt. Add a new canonical register item only after its receipt reopen condition is met, then change status, boundary, or acceptance tests in that JSON:

python scripts/build_deferred_work_register.py
python scripts/validate_closeout_resolutions.py
python scripts/validate_deferred_work.py

Then update this roadmap only when the strategic direction changes. Completed items leave the active register only after their acceptance artifacts exist; the commit history preserves the prior state.