DueCare Project Status¶
Current as of 2026-07-28.
Active Submission Scope¶
The active Kaggle path is exactly three script-kernel folders:
| Kernel | Role | Live status checked 2026-07-28 |
|---|---|---|
kaggle/01-duecare-exploration-workbench/ |
Broad interactive workbench for chat, harness comparison, extraction, search controls, knowledge packs, traces, and activity logs. | COMPLETE |
kaggle/02-live-demo/ |
Focused live demo and video narrative path. | CANCEL_ACKNOWLEDGED |
kaggle/A-00-omni-experiment-workbench/ |
Quantitative proof path: baseline vs harnessed arms, synthetic data, fine-tune, judging, and report artifacts. | CANCEL_ACKNOWLEDGED |
The public A-00 Kaggle page attaches
taylorsamarel/duecare-proof-finetuning-data. That dataset is a guarded
preview, not the full advanced corpus, and Kaggle reports it ready. The latest
A-00 execution is canceled; a fresh successful run and artifact review are
required before it is cited as completed proof. No production adapter is
published.
The interim training collection is public: the SFT and preference dataset views both report ready, and the integrity, CPU training-plan, and four-arm evaluation notebooks all completed on 2026-07-15. These are exact-row companions to the small approved proof release. The completed training starter emitted a plan but did not execute GPU training, so this does not close the adapter or model-lift work.
The retired A-series notebook ladder other than active A-00, task-notebook
snapshots, and older checklist/status docs are archived under
docs/_archive/2026-05-16-legacy-notebook-era/ or
kaggle/_archive/notebooks/. Root kaggle/ should not contain appendix
A-* folders other than A-00-omni-experiment-workbench, and the only root 04-* folder should be
04-kaggle-community-benchmark.
Current Technical Posture¶
- Root
AGENTS.mdnamesmasteras the release branch. The pre-closeout sequence advanced through pull requests 11 through 16. Pull request 16 is the immediate fully merged predecessor to the maintenance closeout at1c8f6b25729da869b2775a29321ab3b74bd4715f; all 16 checks passed. Pull request 15's 4,646-pass run is retained only as older historical evidence. Maintenance mode was enacted on 2026-07-28 with no release tag; every future release candidate must rerun its own gates. - The wrap-up validation made no Ollama, hosted-model, or Kaggle-quota calls.
Render remains the production website/API host; the independent
duecare-ai-siterepository preserves a read-only 51-route continuity copy without a productionCNAME; and this repository's Pages site remains MkDocs. Pull request 14 also fixed the homepage story grid that had squeezed paragraph text into the 28-pixel ordered-list number column. - A later 2026-07-27 live audit corrected the operational boundary: Hermes,
the server-automation vetter, and orchestrator watchdogs/processes were still
active independently of the paused autonomous engine. No local provider ledger existed for that
period, so historical background provider usage is unknown and must be
reconciled privately at the provider. The whole stack is now cost-stopped:
all five recurring tasks are disabled, four stop sentinels are present, and
zero verified repository daemon processes remain. The ignored receipt is
reports/cost_stop_status.json; the public runbook isMAINTAINER_HANDOFF.md. - The repository has a single model-free publication entry point:
PUBLICATION_READINESS.mdandpython scripts/validate_publication_readiness.py --scope core. All twelve current core gates pass in the maintenance candidate; rerun them on any future exact release commit. CLOSEOUT_RESOLUTIONS_2026_07_28.mdrecords the honest outcome and claim boundary for all 11 inherited items.DEFERRED_WORK.mdcontains zero current items and is reserved for specifically reopened work.- All 18 Python distributions remain unpublished on PyPI by explicit closeout decision. One fail-closed OIDC workflow owns package publication; a reviewed independent-SemVer manifest now reconciles the intentionally mixed versions and selects one package per production tag.
- Maintainer succession is now explicit:
CLAUDE_CODE_HANDOFF.mdis the tracked coding-agent pickup,MAINTAINER_HANDOFF.mdis the human operational pickup, andPROJECT_TRANSITION_PLAN.mdpreserves the 2026-07-26 through 2026-08-25 successor playbook. The separate--scope handoffgate is read-only and keeps both documents linked to live pickup evidence. - Public deployment ownership is reconciled:
duecare-ai.comdeploys fromapps/duecare-ai.comon Render; the separateduecare-ai-siterepository deploys the backend-free continuity Pages site; and this repository deploys MkDocs throughdocs-deploy.yml. Its website exporter uploads both artifacts but cannot overwrite MkDocs Pages. The website exposes/project-statusas the public continuity entry point, and the fallback manifest records itssource_revision. - A later owner decision now defines the post-competition endpoint: keep Render
and the current centralized hub available through competition grading, then
retire centralized Fly/Render-style hosting only after the event-triggered
Post-Competition Hosting Transitionpasses. GitHub Pages will preserve the public presentation; mutable hub APIs, submissions, accounts, admin, and automation will no longer be available centrally. The runtime and FastAPI hub remain deployable as independently governed local or partner nodes. - The website's advertised schema URLs now have local route tests instead of pointing at 404s. Kaggle pages show point-in-time run status and distinguish public notebooks from private owner-side drafts.
- Containerization, automation, outreach, and human validation have distinct
boundaries. The runtime and public hub are containerizable, but Hermes,
the server-automation vetter, the orchestrator, and the autonomous engine are
host-scheduled processes and are currently disabled.
scripts/hermes.pystages propose-only synthetic research prompts; it neither contacts civil-society members nor collects ratings. - The public outreach API detects gaps, suggests public support organizations,
matches consented address-hash/topic profiles, drafts campaigns, and vets
manually forwarded observations. It cannot send mail because it stores no
raw recipient addresses. A curator needs a separately owned, consented
address book to resolve profile hashes. The SMTP/IMAP and Hermes-mail architecture in
deployment/oracle_email_solicitation.mdis explicitly a future reference design. - The human-validation packet contains 364 items across 182 strata and has zero qualified independent human ratings. Automated rubric scores and LLM-judge outputs are not human input and do not establish agreement or field effectiveness.
- The industry-neutral system design is now explicit in the Capability-Gap Harness and Network Blueprint. It separates the gap/domain pack, evidence, harness, evaluation, human governance, public-network, and agent-operations planes so the pattern can be replicated outside trafficking without copying trafficking-specific claims. The document also distinguishes a container-friendly target topology from DueCare's current host-scheduled Hermes, server-automation, orchestrator, and autonomous-engine processes.
- Broad local/hosted Ollama work is deferred. The rich harness supports a
non-mutating
--planand a startup ceiling through--max-planned-model-calls/DUECARE_MAX_PLANNED_MODEL_CALLS. - A later owner-authorized Kimi K3 access check on 2026-07-28 used the verified
Ollama catalog ID
kimi-k3and a five-attempt, 20,000-input-token, 3,840-output-token, US\(0.25 ledger cap. All five requests reached Ollama but returned HTTP 402 because the account had no extra-usage balance. The local receipt records zero successes, provider tokens, and actual cost. These are access failures, not Kimi results; the proposed 500-prompt lane was not run. Its frozen no-call plan selects 500 public synthetic prompts across 117 categories (seed20260728, selection SHA-2569d4aedf042f5f9d73e8372a8f1bf5538190d9791dbc692c38ca720aed1bc48eb), reserves 158,922 estimated input and 384,000 maximum output tokens, and has a US\)6.2368 worst-case reservation at the checked rates. - A subsequent paired Kimi K3 smoke used official tag
kimi-k3:cloudon one hash-bound public prompt, once baseline and once with the full DueCare harness. The harness preflight now correctly includes deterministic tools; this prompted a real adapter fix and regression test. One GREP rule, eight RAG documents, and four tool calls were assembled, but both hosted calls returned HTTP 402. The two-attempt receipt reserved 3,491 input tokens, 3,000 output tokens, and US$0.055473, with zero completions, provider tokens, or actual cost. Therefore Kimi quality and Kimi-specific harness lift remain unknown. This is a funded-later next step, not a closeout blocker. The exact public receipt iskimi_k3_harness_lift_smoke_20260728.json. - A subsequent frozen campaign now extends that candidate lane with a local
deterministic grade on every successful answer, 500 cross-family
gemini-3.1-pro-previewcontextual judgments, and 500 separately labeled Kimi K3 contextual self-judgments. The holistic judge protocol is directional, not publication-grade per-dimension adjudication. Across all three hosted phases, the exact maximum is 1,500 calls, 7,296,582 estimated input tokens, 1,152,000 maximum output tokens, and US$34.448916 worst-case. The campaign remainsblocked_external_access: Kimi extra usage is unfunded and no Gemini API credential is present. It has zero candidate completions, automated judgments, or human ratings. The authoritative receipt iskimi_k3_500_context_judge_campaign.json. - Calls entering the primary
llm_generate.pyrouter now also pass through a shared SQLite attempt/token/cash ledger. Offline tests prove zero-call mode blocks before HTTP transport and that retries consume separate reservations. Both the baseline model-failure candidate caller and its contextual judge caller are also covered after the Kimi access check. This does not yet intercept every other direct package/application/ standalone client or self-contained Kaggle kernel; the exact contract isPROVIDER_BUDGETING.md. - All three scheduled model-call wrappers now fail closed unless the shared router has
a positive attempt cap, finite input/output/cash caps, a stable run ID, and
reviewed pricing or an explicit recorded unknown-cost override. Their
watchdog registration and
-Runpaths no longer remove pause sentinels. - The strict training-data audit is not clean: five dense generic-corridor typologies produced 25 metadata-only curation tasks and a 75-row minimum expansion target. A deterministic workbook now defines all 75 risk, benign-neighbour, and counterfactual slots with split-safe lineage, language, perspective, source, and two-person-review requirements. Its 12 official-source candidates remain quarantined and all 75 content slots are honestly unfilled. Existing published learning artifacts remain bounded by their original manifests; this blocks a new training claim, not review of the core repository.
- Local Gemma inference is standardized through
Gemma4Runtime.load(). - Active A-00 uses the same offline default harness as the Kernel 01 comparison path: Persona + GREP + RAG/context + deterministic tools, with internet/import off.
- Active A-00 can run local Gemma judging by default and optional external judge paths when credentials are present.
- Active A-00 training supports checkpoint/resume, adapter save/load, and final report export.
- The harness system is documented as an ecosystem rather than a single monolithic harness.
Current Docs To Trust¶
architecture/capability_gap_blueprint.md- reusable industry-neutral architecture, domain-pack boundary, evaluation triangulation, human network, and agent/container promotion path.research/model_failure_run_readiness.md- exact Kimi/Gemini campaign topology, blockers, plans, and execution contract.kaggle_final_closeout_post.md- exact final Kaggle community update covering new work, honest gaps, and node-first future.POST_COMPETITION_HOSTING_TRANSITION.md- keep Render running through grading, then perform the validated Pages/DNS/data transition without implying that static hosting retains central APIs.CLAUDE_CODE_HANDOFF.md- current tracked coding-agent pickup, public services, recent receipts, boundaries, and exact safe next work.MAINTAINER_HANDOFF.md- fresh-shell operations, system boundaries, access transfer, recovery, and acceptance.PROJECT_TRANSITION_PLAN.md- dated 30-day closeout, successor rehearsal, release decision, and maintenance fallback.FOR_PEER_REVIEW.md- reviewer verification path.PUBLICATION_READINESS.md- offline release gate, Ollama-credit plan, dataset roadmap, and additional source candidates.DEFERRED_WORK.md- canonical unfinished-work register; this supersedes duplicated backlog tables in narrative docs.USER_TODO.md- manual actions before submission.readiness_dashboard.md- current status snapshot.two_week_submission_plan.md- final execution plan.current_kaggle_notebook_state.md- active versus archived Kaggle inventory.harness_ecosystem.md,harness_pattern.md, andharness_standard_contract.md- harness contract and inventory.model_loading_trace.md- Gemma 4 runtime contract.
Remaining Work¶
The canonical queue is DEFERRED_WORK.md. The numbered
summary below is directional only; the register controls status and acceptance.
- Complete the 30-day access, recovery, release/no-release, successor rehearsal, and final acceptance actions in the transition plan.
- Keep the three active Kaggle source surfaces runnable; rerun 02 only when a new recording is needed and A-00 only for a deliberately funded proof.
- Approve source snapshots, then fill and independently adjudicate the 75-row corridor workbook before making a new fine-tuning or adapter-improvement claim.
- If the frozen Kimi/Gemini campaign is funded, re-run all three no-call plans, investigate any selection/context/token drift, then authorize candidate, Gemini cross-family judge, and Kimi self-judge phases separately under the shared ledger. Preserve checkpoint/resume state and keep Meta Muse Spark 1.1 as a required future comparison lane after revalidating its exact provider identifier, access, capabilities, and pricing.
- Attach exact report, activity, prompt/response, training-metadata, chart, and evidence artifacts to any new claim.
- Keep A-00 as the active optional proof path. A-30 and the other appendix
notebooks remain archived under
kaggle/_archive/notebooks/. - When the owner confirms competition grading is complete, publish the final Kaggle closeout post and execute the post-competition hosting transition; until then, keep Render and all current public URLs running.
Verification¶
Use the focused commands in FOR_PEER_REVIEW.md. The docs
contract tests intentionally fail if current entry docs drift back toward the
retired appendix-ladder framing.
Latest pre-handoff offline receipt on 2026-07-28:
- publication handoff: 2/2 gates passed, including 23/23 succession-document checks and 65/65 live pickup checks;
- publication core: 12/12 gates pass, including provider-budget coverage, zero-item deferred-work integrity, the dated 11-item closeout receipt, and the 18-package release-ownership/install-truth gate;
- combined
packages testsregression in the closeout 18-package workspace: 4,669 passed and 9 skipped in 7 minutes 57 seconds with no warning summary under the zero-call transport lock; the focused 43-test package run also passes with pandasRuntimeWarningpromoted to an error; - MkDocs strict build: passed with zero warnings; existing repository-relative source links are resolved to canonical GitHub URLs without suppressing genuinely missing targets;
- training readiness: strict quality and provenance failed as documented,
while the 25-task corridor plan and 75-slot scaffold passed privacy/safety
validation; strict completion remains red at
0/75, as intended; - focused deferred-work/handoff/publication/site tests: 96 passed, including the complete 78-test website suite;
- curator-block governance: all 75 universal rubric dimensions have inline guidance; strict validation reports zero errors and zero warnings;
- post-deploy external network audit: 592 links checked, zero confirmed broken, and nine transient/unverified hosts kept separate from confirmed failures;
- the polish/validation commands initiated no Ollama or hosted-model call; the separate background-daemon usage correction above supersedes any broader reading of that receipt.