30-Day Project Transition Plan¶

This plan turns the final month of primary-maintainer involvement into a verifiable transfer rather than an informal document dump. Operational pickup is defined in the Maintainer handoff; the public release boundary, limitations, dataset plan, and model-work backlog remain in Publication readiness.

Status: active closeout Start date: 2026-07-26 Target handoff date: 2026-08-25 Model-quota posture: locked by default; deterministic work only unless a separate frozen, budgeted run is approved

If the actual departure date changes, move the calendar dates together while preserving the sequence, evidence requirements, and final rehearsal window.

Definition Of Done¶

The project has reached a good stopping point when:

a named successor can establish live state, explain system boundaries, run the handoff and core release gates, and recover from common incidents;
or, if no successor is available, DueCare is deliberately placed in a documented maintenance mode with the engine paused and no implied support or fresh-model promise;
an exact commit/tag and bounded claim set are chosen, or a written no-release decision is recorded;
package versions, CITATION.cff, changelog, release notes, datasets, active Kaggle references, and public docs agree at that exact revision;
access, recovery, domain/hosting, registries, provider billing, outreach, and archive ownership have private transfer receipts;
core, handoff, and relevant focused gates have saved receipts, while any red training or experimental lane is plainly described and isolated;
no raw worker data, private case material, credentials, ignored staging reports, or unreviewed entity allegations are published; and
the successor's first small backlog item has a clear owner, acceptance test, and explicit model-credit budget.

Workstreams And Deliverables¶

Workstream	Required deliverable	Evidence
Current truth	Live-state inventory and deliberate-stops register	Handoff/core receipts, engine status, branch/status snapshot
Product boundary	Maintainer can explain active, optional, archived, and propose-only surfaces	Handoff rehearsal notes and architecture walkthrough
Release	Exact release or no-release decision	Commit/tag, release notes, version/citation reconciliation receipt
Data and research	Accepted datasets separated from candidate, quarantine, and experimental evidence	Manifests, lineage/audit results, limitations register
Operations	Access, recovery, alerts, backups, and publishing authority transferred	Private least-privilege transfer receipts
Public deployment	Render website and MkDocs Pages have distinct, tested ownership	Website health, docs workflow receipt, and absence of a competing Pages deployer
Documentation	Public entry points and purpose maps lead to current handoff/release truth	Handoff validation and public-surface audit
Continuity	Successor completes a fresh-shell pickup, safe change, and restore rehearsal	Signed or dated rehearsal record outside the public repo if it names people

Week 1 - Inventory And Freeze Candidate¶

2026-07-26 through 2026-08-01

Freeze scope: no broad new features, datasets, judging sweeps, or model runs.
Inventory live Git/process/engine state and compare it with saved .claude state, ignored reports, status docs, and public claims.
The reconciled codex/full-flywheel-training-20260714 closeout branch was merged through pull request 2 to master as 07cfdbfd00e1bc304ffc1f8b2736c4d93bbf0eab. Post-merge CI, MkDocs Pages, the artifact-only website build, and the Render project-status route were verified on 2026-07-26. Use current master as the integrated base.
Land this handoff, transition plan, deterministic handoff validator, purpose maps, and navigation links.
Confirm the canonical active Kaggle surfaces and keep optional/archived lanes labeled correctly.
Classify every open item as release-blocking, successor backlog, experimental, or deliberately stopped.
Choose a release candidate claim set. The default is a core/docs/data release with no new training/model-improvement claim while the strict training gate remains red.
Start the private access inventory without copying credentials into tickets, chat logs, or the repository.

Exit evidence: handoff and core gates green in the working environment; known red lanes recorded; candidate release boundary agreed.

Week 2 - Release Candidate And Transfer Package¶

2026-08-02 through 2026-08-08

Reconcile package/workspace versions, CITATION.cff, changelog, dataset versions, Kaggle pins, and release notes as one proposed release decision.
Run focused regressions, the full practical regression suite, public-surface checks, package collection, Kaggle validators, and privacy-safe secret/data scans. Preserve exact warnings and skips.
Build the source/release artifacts reproducibly and record hashes; do not upload them yet unless external publication is separately authorized.
Verify a durable archive and isolated restore, including generated evidence needed to support published claims.
Complete private access-transfer invitations using least privilege, and test recovery/second-factor paths without sharing recovery material in Git.
Prepare a one-hour architecture/evidence walkthrough using one benchmark trace and one dataset-lineage trace end to end.

Exit evidence: release-candidate receipt, artifact hashes, restore receipt, and access-transfer matrix ready for successor rehearsal.

Week 3 - Successor Rehearsal And Knowledge Transfer¶

2026-08-09 through 2026-08-15

Have the successor execute the First 30 Minutes sequence from a fresh shell without relying on the outgoing maintainer's command history.
Rehearse a documentation-only change through review, focused tests, public validation, commit, and rollback/recovery planning without publishing.
Walk active workbench behavior, shared model loading, deterministic/model boundaries, benchmark comparability, dataset admission, entity-intelligence quarantine, and archived Kaggle surfaces.
Rehearse incident responses for unexpected provider traffic, a suspected secret, a worker-data leak, stale generated state, and a failed core gate.
Let the successor identify one confusing or fragile area; improve the docs or test that would have prevented the confusion.
Decide maintenance cadence, issue/security intake, release authority, and the first 30-day successor backlog.

Exit evidence: successor pickup receipt, safe-change rehearsal, incident walkthrough, and documented corrections from rehearsal feedback.

Week 4 - Final Fixes And Release Decision¶

2026-08-16 through 2026-08-22

Fix only issues discovered by rehearsal or release validation that threaten safety, reproducibility, pickup, or the bounded release claim.
Rerun affected tests, handoff scope, core scope, pickup validation, and public surface checks on the proposed final revision.
Verify duecare-ai.com/project-status from Render and the MkDocs handoff pages from GitHub Pages after the release revision reaches master; do not deploy a marketing bundle into the documentation Pages environment.
Record the training gate as passing or explicitly red. Do not weaken it and do not delay a valid no-new-training-claim core release solely to manufacture a green training lane.
Confirm the incomplete per-dimension lane remains isolated from the default comparable board and is described as partial evidence.
Finish private ownership/recovery receipts and remove outgoing access only after incoming access has been tested.
Choose one of three explicit outcomes: publish/tag the bounded release, defer release with a dated reason, or enter maintenance mode.

Exit evidence: final decision record, exact revision, complete validation receipt, limitations, and tested successor or maintenance-mode owner.

Final 72 Hours¶

2026-08-23 through 2026-08-25

Freeze nonessential changes and rerun live-state, handoff, core, and pickup checks on the exact final revision.
Confirm domain/hosting, GitHub, Kaggle, registry, provider, outreach, alerts, billing, backup, and recovery ownership from the private receipt.
Publish/tag only the previously approved bounded artifacts. Record immutable links and hashes without copying secrets or private receipts into the repo.
Verify the public landing page, documentation, release, active Kaggle links, and citation/version text from an unauthenticated view.
Leave the autonomous engine paused. A future maintainer may resume it only through an explicit budgeted plan and current authorization.
Archive the closeout receipts and limitations, then remove outgoing access according to the agreed recovery plan.

No new feature, dataset family, provider integration, or broad benchmark run enters this window.

Owner-Only Actions¶

These actions require the current account owner or an explicitly authorized successor; local repository edits do not complete them:

approve the release/no-release/maintenance-mode decision and bounded claims;
transfer or invite GitHub, Kaggle, hosting/domain/DNS, PyPI, Hugging Face, monitoring, archive, shared-mailbox, and model-provider access;
verify billing, recovery, two-factor, security-contact, and revocation paths;
rotate any credential involved in an incident and review platform audit logs;
publish a GitHub release, package, model/dataset, Kaggle revision, website deployment, or competition submission;
authorize any nonzero model-credit budget or removal of the engine stop sentinel; and
decide whether to retain, transfer, or retire public support commitments.

The repository may document commands and public identifiers, but must never store the private receipt's secrets, recovery answers, personal contact details, or billing information.

Decision Register¶

Decision	Current disposition	Revisit trigger
Core release versus new training claim	A bounded core release may proceed while training stays red, provided it explicitly makes no new training/model-improvement claim	Strict quality and provenance gates pass on a new append-only record
Model usage during closeout	Zero planned calls by default; no Ollama or hosted-model spend is needed for closeout	Owner approves a frozen sampled plan with finite allowance and stop condition
Autonomous engine	Intentionally paused; no automatic resume	Explicit current authorization plus live preflight, budget, and review plan
Comparable benchmark board	Keep v1/h1 batched evidence as default	A versioned successor board is complete and independently documented
Exhaustive per-dimension lane	Generation complete, judging incomplete, experimental and isolated	Exact coverage manifest reaches closure under a frozen budget
Entity-intelligence pipeline	Propose-only, curator-reviewed, separate from worker-facing and training paths	A reviewed governance and evidence-admission change is approved
Kaggle surface inventory	`01`, `02`, and `A-00` active; `03` and `04` optional; notebook-era variants archived	Root `AGENTS.md` and `kaggle/_INDEX.md` change together
Release version/tag	Pending a deliberate owner decision	Week 2 release-candidate reconciliation is reviewed
Target branch	Root rules name `master`; pull request 2 merged the reconciled closeout as `07cfdbfd00e1bc304ffc1f8b2736c4d93bbf0eab`, and post-merge CI plus both public deployment paths were verified	Continue from current `master`; rerun gates on the exact candidate before any tag
Successor	Transfer path preferred; maintenance mode is the safe fallback	A successor accepts and completes the rehearsal

Exit Criteria¶

Before declaring the transition complete, retain evidence that:

If No Successor Is Available¶

Use maintenance mode rather than leaving an ambiguous live project:

Keep the engine paused and the zero-call default documented.
Freeze the last supported commit, release artifacts, checksums, docs, and limitations; do not imply that volatile legal/operational facts stay fresh.
Add a dated public maintenance notice that names the support and security posture without exposing personal contact information.
Disable or narrow unattended deployments, provider credentials, write integrations, webhooks, and billing only through their owner-authorized platform controls.
Preserve the repository and public research artifacts unless the owner makes a separate archival decision. Repository archival is an external/manual action, not an automatic closeout step.
Retain a private recovery owner for domain, hosting, security reports, and backups, with a periodic renewal/expiry calendar.
Keep candidate data, unreviewed entity matches, partial judge lanes, and ignored reports out of publication.

Maintenance mode is a responsible stopping point, not a claim that every research lane or product roadmap item is complete.

Future Improvements¶

These are ordered for a successor; they are not exit blockers unless promoted by the Decision Register.

Reliability and maintenance¶

Add the handoff/core gates to continuous integration with pinned Python versions and artifact receipts.
Automate version/citation/changelog consistency checks and reproducible release builds without automating publication authority.
Isolate legacy Ruff cleanup into small behavior-preserving changes.
Clear the 84 classified legacy strict-MkDocs warnings while preserving the intentional public-versus-provenance navigation boundary.
Refresh GitHub Action majors after checking the current official releases; the green Pages receipt still emits a Node.js 20 runtime deprecation notice.
Add restore drills, dependency/security review cadence, and a dated legal and public-source freshness dashboard.

Ollama and provider calls¶

Put every provider attempt behind one atomic ledger that reserves and settles calls, input/output tokens, estimated cash, retries, and cancellation.
Make retry classification explicit: retry transient transport/service errors only; fail closed on authentication, permission, quota, and invalid requests.
Use content-addressed generation/judging caches keyed by provider, immutable model revision, prompt, rubric, decoding settings, and harness version.
Add preflight capability discovery and a dry-run plan that cannot start a daemon, write a heartbeat, or mutate result artifacts.
Spend first on a small stratified smoke matrix, then allocate extra judging only to disagreement, calibration, and decision-boundary cells.
Emit sanitized budget and coverage receipts so an interrupted run can resume without duplicate paid calls.

Dataset and evaluation quality¶

Close the current 25-task / minimum 75-row corridor-diversification plan with lawful dated sources, lineage, two-person adjudication for severe cases, and no threshold weakening.
Add dataset cards with license/terms snapshots, inclusion/exclusion reasons, transformation code, source freshness, language/corridor coverage, and known harms for every publishable dataset family.
Strengthen near-duplicate and lineage-family leakage detection across supervised, preference, reward, benchmark, quarantine, and held-out splits.
Add multilingual and code-switch evaluation designed with native speakers, including abstention, translation drift, and culturally specific hard negatives.
Build temporal legal-freshness and source-ablation lanes so the project can measure when retrieval is stale or a response depends on one weak source.
Publish human-adjudication protocol, agreement, disagreement taxonomy, severity stratification, and benign-control performance before stronger safety claims.
Keep synthetic/model-generated candidates labeled and quarantined until provenance, privacy, diversity, and human/rule-based admission gates pass.

Research and product evidence¶

Complete the isolated per-dimension lane only under a frozen budget and exact coverage manifest; never backfill it into v1/h1 silently.
Add calibration, abstention, uncertainty, and selective-judge escalation reporting rather than relying on aggregate score lift alone.
Run a small prospective study with NGO/caseworker review, ethics and privacy controls, predeclared outcomes, and a clear boundary between usability and field-effectiveness claims.
Test cross-corridor counterfactual consistency and subgroup error patterns, then publish harms and regressions alongside wins.
Convert repeated operational lessons into narrow tests and decision records before expanding the application surface.

The best first successor milestone is intentionally modest: keep all model calls locked, make the handoff/core gates reproducible in a fresh environment, close one evidence-backed documentation or dataset-quality gap, and preserve the project's honest claim boundary.