Publication Readiness And Next Work

This is the model-free wrap-up path for DueCare. It separates the maintained public surface from explicitly excluded claims and work that may be reopened later only under a dated condition. Current as of 2026-07-28.

For the transferable system design behind this domain implementation, start with the Capability-Gap Harness and Network Blueprint.

Current Posture

Track	Current state	Publication rule
Core code, docs, active Kaggle surfaces, provider-budget coverage, deferred-work and closeout-receipt integrity, package-release ownership, and published dataset claims	12/12 current core gates pass offline; the 2026-07-28 maintenance candidate passed 4,669 tests with 9 skips locally, while pull request 16 at 1c8f6b25 is the immediate fully merged predecessor	Maintenance mode is enacted; rerun on any future exact release commit before tagging.
Maintainer succession and live pickup	2/2 handoff gates passed offline; the tracked Claude handoff, Render site, independent read-only continuity Pages site, and MkDocs docs are reconciled	Re-run from a fresh shell and complete the private/manual acceptance steps before ownership transfer.
Public hosting transition	Event-gated: Render and duecare-ai.com stay live through competition grading; afterward Pages is the approved durable presentation target and independently governed nodes remain deployable	Do not shut down, change DNS, or remove secrets/data early. Follow every acceptance item in POST_COMPETITION_HOSTING_TRANSITION.md; Pages does not preserve mutable hub APIs.
Existing dated benchmark and learning-study results	Retained as bounded evidence with their original model, dataset, rubric, and date	Do not silently relabel an old result as a new model or field-effectiveness result.
New fine-tuning dataset	Not clean yet: the strict audit reports five dense single/generic-corridor typologies; the deterministic workbook has 75 unfilled slots and no fabricated approvals	Complete source snapshots, rights review, two-person adjudication, and lineage-safe rows; rerun the audit, then refresh provenance before training or a new model claim.
Exhaustive per-dimension judging	Experimental, isolated, and incomplete	Keep it out of the default comparable board until its own exact closure gate passes.
Kimi/Gemini directional campaign	Frozen but externally blocked: 1,500-call maximum; zero candidate completions, automated judgments, or human ratings; Kimi extra usage is unfunded and no Gemini key is present	Re-run the no-call plans and authorize candidate, Gemini cross-family judge, and Kimi self-judge phases separately; never blend the self-judge into the primary automated result.
Other local/hosted model work	Optional and deferred; the whole Windows model/flywheel stack is cost-stopped and seven registered transports have atomic attempt/token/cash coverage	Plan offline first; unlock a small allowance only for a frozen, priced run, and keep uncovered direct/notebook clients explicitly labeled.

The dated CLOSEOUT_RESOLUTIONS_2026_07_28.md receipt is authoritative for the 11 inherited decisions. The generated DEFERRED_WORK.md register contains zero current items and is reserved for work whose receipt reopen condition is actually met.

Registry And Kaggle Publication Truth

• All 18 duecare-llm* distributions are buildable from source, but none had a public PyPI project on 2026-07-27. docs/PACKAGE_INVENTORY.md is the canonical install/version map.
• .github/workflows/pypi-publish.yml is the sole publisher. Production PyPI has no manual dispatch target and no generic v* trigger. A package-NAME-vMAJOR.MINOR.PATCH tag must match exactly one row in the reviewed independent-SemVer manifest. The current 0.1.0 / 0.1.2 / 0.17.0 mix is intentional and no longer a policy blocker.
• Live Kaggle status checked 2026-07-28: active 01 is COMPLETE; active 02 and A-00 are CANCEL_ACKNOWLEDGED; optional 04 is COMPLETE; optional 03 has no verified public URL. A canceled run is not completion evidence.
• No additional notebook is required for repository closure. The private and built queue in NOTEBOOKS.md should advance one item at a time only when it closes a named audience or evidence gap.

The quality audit currently has zero SFT split leaks, zero DPO split leaks, zero incoherent citations, and zero phone-like hits in gold text. Its blocker is coverage shape: five typologies have enough rows to learn a generic-corridor shortcut. The resulting curation plan contains 25 metadata-only tasks in five batches and recommends at least 75 reviewed rows. The queue contains no raw prompt or answer text and passes its privacy checks.

The deterministic curation workbook now expands those tasks into exactly 75 review slots: 25 risk cases, 25 benign near-neighbours, and 25 corridor counterfactuals. It balances six perspectives, plans 45/15/15 rows across train/validation/test by whole lineage family, and assigns English plus relevant Bangla, Malay, Arabic, Amharic, and Hindi review lanes. The source catalog has 12 official or intergovernmental candidates, but every one is deliberately blocked from training until a compatible-rights decision, immutable snapshot, retrieval date, and SHA-256 exist. No source or row is represented as human approved.

Refreshing the audit and corridor plan invalidates fingerprints in the older append-only planned fine-tune record. That is expected: do not rewrite history or append a replacement record until the quality audit is clean. The strict training-provenance gate should remain red in the meantime.

One-Command Offline Review

Run from the repository root:

python scripts/validate_publication_readiness.py --scope core

This runs public-surface and messaging audits, the source-checkout harness smoke test, published dataset-claim verification, fallback-registry validation, both active Kaggle static gates, the generated deferred-work register, the dated closeout-resolution receipt, package-release reconciliation, and package test collection. The runner sets Ollama's planned-call allowance to zero and forces common Hugging Face and Weights & Biases integrations offline for its child checks.

Current audit receipt: all twelve current core gates pass in the model-free maintenance candidate without a model or network call. Its immediate fully merged predecessor landed through pull request 16 as 1c8f6b25729da869b2775a29321ab3b74bd4715f; all 16 checks passed. Pull request 15's exact local broad regression passed 4,646 tests with 9 skips, and the subsequent tracked-handoff candidate passed 4,648 tests with 9 skips; both are historical. The current maintenance candidate passed 4,669 tests with 9 skips in 7 minutes 57 seconds under the zero-call lock and offline provider/model flags. These receipts are not a substitute for rerunning the command on an eventual release tag.

Training readiness is intentionally separate:

python scripts/validate_publication_readiness.py --scope training

It first requires all 75 source-bound and independently adjudicated rows, then runs the strict quality audit, validates the corridor plan, and verifies the fine-tune registry/model-card/trainer provenance chain. A nonzero result is the honest maintenance state. It changes only if a future, independently reviewed corridor dataset is deliberately admitted.

For local handoff state and live automation status, also run:

python scripts/validate_project_bible_pickup.py

That command is workspace-specific evidence; it is not part of the portable publication gate.

Transition And Succession

This file remains the release and evidence boundary. Coding agents start with the tracked Claude Code handoff. For human ownership transfer, use the operational Maintainer handoff and the dated 30-day transition plan. Their read-only gate is:

python scripts/validate_publication_readiness.py --scope handoff

That scope validates succession structure, discovery and local links, category/count-only privacy checks, public website/Pages ownership, the public continuity route, and live Project Bible pickup consistency. It does not make the strict training lane portable or green, and it does not authorize publication, credential transfer, provider spend, or engine resume. The current working-tree receipt is 2/2 handoff gates passed, including 23/23 succession-document checks and the 65-check pickup validator with zero findings.

Preserve Ollama Credits

Keep the current shell locked while doing deterministic work:

$env:DUECARE_MAX_PLANNED_MODEL_CALLS='0'

On the Windows automation host, also require the whole-stack receipt:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File scripts/stop_ollama_stack.ps1 -Status

The 2026-07-27 live audit found that the discovery and server-automation daemons had continued independently after the autonomous engine pause. No repo-local provider ledger covered that background period, so its historical usage is unknown and must be reconciled privately at the provider. The current status requires all five recurring tasks disabled, all four daemon sentinels present, and zero verified repository daemon processes.

For the rich harness, always inspect the exact incremental plan before a run:

python scripts/rich_harness_lift.py --n 40 --plan --require-complete

--plan now exits before heartbeat or result files are written. The rich harness also accepts --max-planned-model-calls N; it exits with code 4 before any model call or run-artifact write when the offline estimate exceeds N. The command-line value overrides the environment lock.

The rich-harness guard counts planned logical generation, judging, and pairwise calls. The primary llm_generate.py router now adds a separate atomic transport ledger: it reserves attempts, estimated input tokens, maximum output tokens, and reviewed worst-case cost before each HTTP request. Retries, key rotations, and resilient re-questions consume new reservations. The full operator contract, positive-budget environment, privacy boundary, and current coverage matrix are in Provider budgeting.

The transport lock is exact for seven registered HTTP paths: four primary router transports, the optional adverse-media verifier, the model-failure candidate client, and the contextual judge client, as enforced by validate_provider_budget_coverage.py. It is not a repository-wide network interceptor: self-contained Kaggle kernels, package/application adapters, and other standalone scripts with their own HTTP clients remain operator-controlled. Keep provider credentials unavailable to those paths during deterministic maintenance.

When model work resumes, use this order:

1. Freeze the prompt-set hash, model IDs/revisions, harness/rubric versions, shuffle seed, maximum output tokens, and expected artifact paths.
2. Run deterministic gates and --plan first.
3. Configure the finite shared provider ledger, then unlock only the sampled allowance and exact caller needed for a small stratified smoke run.
4. Reuse successful generation and component checkpoints. Never discard a resumable cache merely to restart cleanly.
5. Retry only transient timeouts or service failures. Do not retry quota, authentication, permission, or invalid-request failures.
6. Escalate judge disagreements or borderline cases to extra judges; do not spend a full panel on every easy cell.
7. Keep v2/h2/per-dimension experiments in versioned files, separate from the v1/h1 batched board.
8. Include Kimi K3 and Meta Muse Spark 1.1 as required comparison lanes. Reverify immutable provider IDs, access, modalities, context, and pricing at run time; preserve an unavailable-lane receipt rather than substituting a different model silently.

The current Kimi campaign is already preregistered in kimi_k3_500_context_judge_campaign.json: 500 Kimi baseline answers, deterministic grades, 500 Gemini 3.1 Pro cross-family contextual judgments, and 500 separately labeled Kimi contextual self-judgments. Its holistic protocol is directional. 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. It has not run. Follow the readiness receipt; access failures are not benchmark rows.

The shared primary-router ledger and three direct-client migrations are complete and tested without provider calls. The remaining transport work is to migrate other direct standalone research clients one at a time, then design a portable notebook equivalent and an injectable package/application-adapter contract without changing existing benchmark evidence lanes.

Dataset Improvement Plan

The current curation plan applies each of five under-diversified typologies to five proposed corridors. At three reviewed rows per task, that is 25 tasks and 75 new rows.

Expansion corridor	Why it is useful
Bangladesh to Malaysia	Recruitment-fee, manufacturing, and intermediary variation
Bangladesh to Saudi Arabia	Domestic-work and Gulf recruitment-chain variation
Ethiopia to Gulf destinations, maritime	Origin, route, and sector variation outside the dominant examples
Ghana to Qatar	West African origin and construction/service-chain variation
India to Kuwait	South Asian origin with a distinct destination-law context

The row review should add diversity without creating a new shortcut:

• split held-out data by lineage/mechanism family, not random row alone;
• balance worker, recruiter, employer, platform, investigator, and bystander perspectives;
• include benign near-neighbours and corridor-swapped counterfactuals to test over-refusal;
• cover relevant languages and code-switching, with bilingual human review;
• keep statute, fee, hotline, office, and URL claims in dated knowledge objects;
• require two-person adjudication for disputed labels and high-severity gold answers;
• attach source URL, retrieval date, license/terms, checksum, transformation, parent hash, and split family to every admitted source-derived row;
• publish a data card and machine-readable metadata alongside row manifests;
• keep private complaints, contact lists, identity documents, and raw case narratives out of the training corpus.

The executable handoff is:

python scripts/build_corridor_curation_workbook.py --validate
python scripts/build_corridor_curation_workbook.py
python scripts/validate_corridor_curation.py
python scripts/validate_corridor_curation.py --require-complete

The first three commands currently pass the metadata-only scaffold. The final command correctly fails with 0/75 valid rows and 75 missing slots. Candidate content belongs only in ignored reports/training/corridor_curation_rows.jsonl. The validator rejects unapproved or checksum-mismatched sources, missing or duplicate reviewers, unresolved disagreements, absent native-language attestation, PII-like text, exact duplicates, cross-family near-duplicates, and lineage-family split leakage. Findings contain row identifiers and category/count summaries, never matched sensitive payloads.

After curation, run these in order:

python scripts/audit_training_quality.py --require-clean
python scripts/build_corridor_expansion_plan.py
python scripts/validate_training_provenance.py --json

Only after a clean audit should the normal training engine append a new provenance record. Do not edit an older append-only registry entry to make its hashes look current.

Additional Public Resources

These are candidate inputs for curated knowledge objects, context features, or evaluation strata. They are not an instruction to bulk-ingest everything. Links were verified against the official publishers on 2026-07-26.

Resource	Best use in DueCare	Admission note
ILO NATLEX	National labour, social-security, and related human-rights legislation	Snapshot the exact instrument and date; a database match is not legal advice.
ILO SDG 10.7.1 recruitment-cost resources	Recruitment-cost definitions, survey design, and fee-burden evaluation features	Keep country/sample limitations with each statistic.
UNODC Data Portal	Aggregate trafficking-in-persons indicators and evaluation context	Use metadata and aggregate strata, not inferred case labels.
IOM Displacement Tracking Matrix API	Aggregated mobility and displacement context	Treat as contextual administrative data; review API terms and geographic granularity.
World Bank WDR 2023 migration data	Bilateral migration matrices and country-level context	Record the original table/figure source and transformation.
U.S. DOL ILAB List of Goods	Country-good forced/child-labour risk strata and hard-negative design	It is a risk source, not proof about a particular company or worker.
Global Fishing Watch vessel-identity data	Vessel identity and maritime-sector entity linkage	Preserve dataset version, license, identity confidence, and match rationale.
Open Supply Hub API	Facility matching, sector/facility context, and supply-chain entity resolution	API access may require a trial/subscription and token; review terms before acquisition.
Open Ownership BODS analysis tools	Beneficial-ownership entity linkage in the propose-only intelligence pipeline	Preserve BODS source, jurisdiction, publication date, and open-license attribution.

Every source must pass the same admission checklist: lawful access, compatible license/terms, minimum necessary fields, privacy review, source snapshot, retrieved_at, checksum, transformation record, and curator approval. External entity data belongs in the propose-only entity-intelligence pipeline until reviewed; it must not flow directly into worker-facing answers or training rows.

The machine-readable candidate registry is configs/duecare/training/corridor_curation_sources.json. It adds corridor-relevant official discovery points for BMET and its live recruiting-agent registry, Malaysia's labour-policy publications, India MEA's overseas-employment material, Saudi HRSD's labour-reform guide, and dated ILO FAIR/FAIRWAY, STREAM, Qatar, and Kuwait material. URL verification is not license approval: all entries remain candidate-only and training-blocked.

Pickup In Ten Minutes

A new maintainer or agent should use this order. None of these commands calls a model:

git status --short
python scripts/validate_publication_readiness.py --scope handoff
python scripts/validate_project_bible_pickup.py
python scripts/autonomous_engine.py --status
python scripts/validate_publication_readiness.py --scope core

Then read:

1. AGENTS.md for safety, active surfaces, and required gates.
2. CLAUDE_CODE_HANDOFF.md for the current tracked pickup, public services, exact receipts, and safe next work.
3. MAINTAINER_HANDOFF.md for operations, boundaries, access transfer, recovery, and acceptance.
4. PROJECT_TRANSITION_PLAN.md for the dated closeout sequence and maintenance-mode fallback.
5. DEFERRED_WORK.md for the canonical unfinished-work queue, authorization boundaries, and acceptance gates.
6. This file for the release boundary and limitations.
7. project_status.md for the concise active-surface snapshot.
8. kaggle/_INDEX.md for active, optional, and archived notebook surfaces.
9. codex/PROJECT_BIBLE.md for the deep historical and autonomous-engine handoff.

The authoritative live/generated evidence is:

Question	Artifact or command
Is the public repository coherent?	validate_publication_readiness.py --scope core
What work is currently outstanding?	The zero-item DEFERRED_WORK.md plus the dated closeout receipt and both validators
Is local automation paused and internally coherent?	validate_project_bible_pickup.py plus autonomous_engine.py --status
Is the new training dataset safe to advance?	reports/training/quality_audit.json and --scope training
What must curators add?	reports/training/corridor_curation_workbook.json, the candidate-only source catalog, and validate_corridor_curation.py --require-complete
Is the exhaustive judge lane complete?	reports/rich_lift/panel_perdim.coverage.json
What Kaggle surfaces are active?	kaggle/_INDEX.md and the two Kaggle static validators

Generated reports under reports/ are generally local/ignored evidence. Do not assume a saved report is current: compare its hashes/timestamps to live status and rerun the read-only validator when possible.

System Map For The Next Maintainer

The map below is the project-specific summary. The reusable seven-plane architecture, generic domain-pack contract, agent roles, human-network loop, and container target are in the Capability-Gap Harness and Network Blueprint.

The three major flows intentionally remain separate:

worker/reviewer input
  -> workbench + shared model service
  -> deterministic GREP / RAG / tools / privacy boundary
  -> optional Gemma generation
  -> versioned evaluation + replay/export evidence

benchmark responses
  -> resumable generation cache
  -> isolated batched or per-dimension judging
  -> exact coverage manifest
  -> dated reports / comparable board

curated data
  -> organize + lineage split
  -> strict quality audit
  -> corridor curation plan
  -> 75-slot source/review workbook
  -> optional train/evaluate
  -> append-only fine-tune registry + verified model card

The entity-intelligence pipeline is a fourth, propose-only flow. It can stage public registry/entity matches for curator review, but it must not feed worker-facing answers, benchmark labels, or training rows automatically.

Deliberate Stops And Known Non-Blockers

• The whole model/flywheel stack is intentionally cost-stopped. A stale engine lock is normal while its sentinel exists; do not remove any sentinel, re-enable a task, or run a daemon merely to make status look cleaner. Use scripts/stop_ollama_stack.ps1 -Status for host truth.
• Generation for the exhaustive per-dimension lane is complete, but judging is incomplete. Its partial coverage is useful research evidence, not default board closure and not a core-publication blocker.
• The training gate is red because of corridor coverage and the resulting stale fingerprints in an older planned registry record. This is a truthful provenance stop, not a reason to rewrite the ledger.
• The current integrated packages tests regression passed 4,669 tests with nine skips and no warning summary in the locked 18-package workspace. Mocked and loopback provider tests explicitly isolate their fake transports while the real zero-call denial tests remain enabled. The focused package follow-up also passes 43 tests with RuntimeWarning promoted to an error, proving the former constant-value pandas Styler warnings are removed rather than hidden.
• The registered legacy Ruff slice is complete: rich_harness_lift.py, verify.py, and test_plan.py pass the agreed lint scope without file-wide suppression. The retired cleanup item must not be reintroduced unless a new, evidenced regression appears.
• MkDocs now passes mkdocs build --clean --strict with zero warnings. The tested repository-link hook rewrites only existing targets outside docs/ to canonical GitHub source URLs; it leaves missing targets untouched so the strict lane remains a meaningful broken-link guard. Informational notices for intentionally unlisted/excluded provenance pages are not release blockers.
• The deployment boundary is now unambiguous: Render owns production duecare-ai.com; the separate TaylorAmarelTech/duecare-ai-site repository owns a backend-free read-only Pages continuity copy with no production CNAME; and docs-deploy.yml owns this repository's MkDocs Pages site. duecare-site-build.yml uploads both source artifacts but never deploys over MkDocs. The continuity manifest exposes its checked source_revision.
• That current ownership has a documented event boundary rather than an immediate change: keep Render active through competition grading. After the owner confirms grading is complete, follow the post-competition hosting transition to make Pages the durable presentation, explicitly retire centralized APIs, and preserve the runtime as independently deployable nodes.
• GitHub Action majors were verified against official releases on 2026-07-27 and refreshed across CI, Pages, website artifacts, scheduled work, Docker, Helm, Gitleaks, and package publication. The triggered CI, Pages, website-artifact, harness-contract, and evaluation jobs passed with the refreshed actions and no Node 20 runtime annotation. Docker, Helm, and PyPI publishing remain release-triggered and should be validated on an approved release candidate/tag, not dispatched as a production smoke test.
• Workspace package versions were not guessed or bumped. ADR-001 now adopts independent SemVer, and configs/duecare/package_release.toml reconciles the intentionally mixed versions with the sole publisher. CITATION.cff still describes living research software because no registry release exists yet.
• The post-deploy concurrent external audit checked 592 outbound links on 2026-07-27 with zero confirmed broken links. All six same-site schema URLs returned 200; nine additional hosts were transient, DNS/SSL-blocked, redirect-looped, or bot-blocked and remain explicitly unverified rather than mislabeled as broken.
• Root AGENTS.md names master as the active release branch. Pull request 16 is the immediate fully merged predecessor to this maintenance closeout at 1c8f6b25729da869b2775a29321ab3b74bd4715f; all 16 checks passed. Pull request 15's 4,646-pass run is older historical evidence. Render, the independent continuity site, and MkDocs retain distinct ownership. A release tag/version remains a separate owner decision.
• Archived notebook-era surfaces are provenance. Do not restore them to the Kaggle root to satisfy old references; update the reference or archive map.
• Never treat a dirty working tree as disposable. Inspect and preserve unrelated edits instead of trying to manufacture a clean status.

Canonical Decisions And Deferred Work

CLOSEOUT_RESOLUTIONS_2026_07_28.md replaces the duplicated 11-item backlog table that previously lived here. It distinguishes completed maintenance and decisions from declined work, claim exclusions, retained current-owner authority, and unknown historical provider usage. DEFERRED_WORK.md now has zero current items. Both artifacts are checked by the core release gate, the handoff validator, focused tests, and CI.

Good research extensions after the release boundary is stable include a cross-corridor counterfactual benchmark, temporal legal-freshness tests, language/code-switch calibration, source-ablation experiments, selective judge escalation, and a small prospective NGO reviewer study. Each should start as a new versioned evidence lane rather than changing the existing board in place.

Release Checklist

• --scope core passes on the exact release commit.
• The release notes name the exact commit/tag, package versions, datasets, rubrics, harness versions, and dated evidence artifacts.
• --scope training either passes or the release clearly states that no new training/model-improvement claim is being made.
• The incomplete per-dimension lane remains labeled experimental and isolated from the default board.
• Secret and sensitive-data scans report categories/counts without printing matched payloads.
• Generated manifests, purpose maps, project status, and handoff artifacts agree with the code, deferred-work register, and current active Kaggle inventory.
• Public copy says which hosting phase is current and does not imply that a static Pages site provides Render's mutable APIs.
• A deliberate release-version decision reconciles workspace package versions, changelog, tag, and the currently unversioned CITATION.cff; do not bump them implicitly during cleanup.

Conditional Future Sequence

There is no current closeout backlog. If a dated receipt reopen condition is met, use this dependency order without treating it as standing authorization:

1. Obtain compatible source rights and immutable snapshots before any independently adjudicated corridor row is admitted.
2. Refresh strict quality and append-only provenance only after genuinely new admitted data exists; do not refresh metadata to disguise a red gate.
3. Run the frozen Kimi/Gemini campaign only after its external access blockers clear and each phase has exact identifiers and finite attempt/token/cash caps. Treat Meta Muse Spark 1.1 or another provider as a separate frozen extension, not a silent substitution.
4. Prefer a small uncertainty-reducing judge or human-review slice over resuming the declined exhaustive per-dimension sweep.
5. Publish a package, notebook, model, or dataset only when a real consumer or evidence need has a maintainer, exact revision, support boundary, and passing gates.