Project continuity · snapshot 2026-07-26

A stable stopping point with an explicit path forward.

DueCare is in a time-boxed maintainer transition through 2026-08-25. Core code, public documentation, active Kaggle surfaces, and published dataset claims have a green model-free release gate. New model spending and a new training-improvement claim remain deliberately paused until their separate evidence gates are ready.

01 · What is ready

Release-ready and research-incomplete are kept separate.

A green core gate means the public repository is coherent at the checked revision. It does not silently turn an unfinished dataset, adapter, or experimental judge lane into a completed scientific claim.

Stable

Core public surface

Public messaging, source-checkout smoke, manifest-backed dataset counts, model-fallback registry, active Kaggle kernels, generated page sources, and package-test collection pass without a model or network call.

Preserved

Dated benchmark evidence

Existing results retain their original model, prompt set, rubric, judge, and date. They are benchmark response-quality evidence, not field-effectiveness, victim-identification, or production-reliability claims.

Paused

Model and engine work

The autonomous engine remains stopped behind its sentinel. Closeout uses a zero planned-call guard; restarting Ollama or a hosted provider requires a new frozen budget, preflight, and explicit authorization.

The strict training lane is not ready.Five dense generic-corridor typologies require a metadata-only 25-task, minimum-75-row curation expansion. The expansion plan passes its privacy checks, but quality and provenance must remain red until reviewed rows, fingerprints, and a verified model card are regenerated through the normal append-only path.
02 · Public deployment ownership

One source and one owner for each public surface.

The website and documentation are related but intentionally deployed by different systems. This prevents one workflow from overwriting the other.

Website

duecare-ai.com

The FastAPI hub in apps/duecare-ai.com deploys from master on Render. It serves product pages, public coordination APIs, sanitized proposal intake, and health endpoints. It does not load Gemma.

Documentation

GitHub Pages

The MkDocs site deploys from docs/ through docs-deploy.yml after changes reach master. It owns the repository's single GitHub Pages site.

Portable build

Static website artifact

duecare-site-build.yml exports the website as a downloadable artifact for a separate root-domain host. It never deploys into the MkDocs Pages environment.

03 · Pick up the project

Start with live evidence, then read the history.

Saved handoff artifacts and ignored reports are historical evidence. A maintainer establishes current truth from Git, live process/engine status, and read-only validators before changing or publishing anything.

git status --short
$env:DUECARE_MAX_PLANNED_MODEL_CALLS='0'
python scripts/validate_publication_readiness.py --scope handoff
python scripts/validate_publication_readiness.py --scope core
python scripts/autonomous_engine.py --status
04 · What remains

The next work is explicit, bounded, and optional by lane.

P0 · Release

Reconcile and freeze

  • Integrate the reviewed branch into master.
  • Choose version, citation, changelog, and tag together.
  • Rerun core and handoff gates on the exact release revision.
P1 · Data

Close the curation queue

  • Add lawful, dated corridor sources.
  • Adjudicate the minimum 75 diverse rows.
  • Regenerate quality and provenance sequentially.
P1 · Providers

Budget every attempt

  • Add one atomic call/token/cash ledger.
  • Classify retryable failures explicitly.
  • Resume with a small cached smoke matrix only.