File Purpose Guide¶
This guide keeps the repository readable for reviewers and future agents. It does not list every generated artifact one by one; instead it defines where each durable file belongs and which index must explain it.
Principle¶
Every committed file should answer one question quickly: "Why does this exist?"
For new durable files, add one of the following:
- A clear title plus purpose paragraph inside the file.
- A row in the nearest directory README or index.
- A link from a higher-level map when the file is a public or judging surface.
Generated reports, scratch exports, cache files, notebook byproducts, and one-off local artifacts should stay out of commits unless they are deliberately published evidence.
Public Purpose Maps¶
| Scope | Purpose map |
|---|---|
| Repository top level | README.md, ROOT_FILES.md, docs/REPO_LAYOUT.md, this guide |
| Kaggle kernels | kaggle/_INDEX.md, kaggle/README.md, kaggle/NOTEBOOK_PURPOSE_AND_RUNBOOK.md |
| Public website | apps/duecare-ai.com/README.md, /project-status, page templates, route and static-export tests |
| Python packages | Each package README.md, package tests, generated component docs |
| Agent handoff | AGENTS.md, CLAUDE.md, PROJECT_BIBLE.md, Plans.md, .claude/rules/ |
| Maintainer succession | docs/MAINTAINER_HANDOFF.md, docs/PROJECT_TRANSITION_PLAN.md |
| Competition docs | docs/FOR_KAGGLE_JUDGES.md, docs/kaggle_writeup_paste_ready.md, docs/kaggle_post_networked_knowledge_sharing.md, docs/video_script.md |
| Publication and peer review | docs/PUBLICATION_READINESS.md, docs/FOR_PEER_REVIEW.md, docs/reproducibility.md |
Generated Module Metadata¶
Many package submodules contain generated PURPOSE.md, AGENTS.md,
INPUTS_OUTPUTS.md, HIERARCHY.md, DIAGRAM.md, TESTS.md, and
STATUS.md files. These are deliberate folder-per-module metadata files. A
STATUS.md file may include a TODO section even when the module has shipped;
that section tracks follow-up completion criteria, not a public placeholder.
Do not delete these files for cosmetic cleanup. If a generated metadata file is wrong, update the source descriptor or the generator so the next refresh keeps the correction.
Repository Root Policy¶
Keep the repository root for GitHub-standard files, tool configuration,
deployment contracts, attribution/provenance files, and the shortest possible
entry points. ROOT_FILES.md is the manifest for the files that intentionally
remain there.
Do not add one-off scripts to the root. Active helpers belong in scripts/;
historical helpers belong under _archive/. Do not add new narrative markdown
files to the root unless they are a GitHub community-health convention or are
added to ROOT_FILES.md with a durable reason.
Durable training helpers in scripts/ must be discoverable from
docs/training_and_finetuning.md. For example,
scripts/validate_publication_readiness.py is the model-free release wrapper:
its core scope composes the portable public, claim, Kaggle, and collection
gates; its handoff scope composes succession and live-pickup checks; and its
separate training scope preserves an honest red result when strict dataset or
provenance checks are not yet clean. scripts/validate_maintainer_handoff.py
checks the two current succession documents, their discovery/local links, and
privacy-safe content without calling a model or network service.
scripts/ollama_adversarial_flywheel.py is the local Ollama candidate producer:
it may generate SFT/DPO candidates and quarantine metadata, but its manifest
must remain candidate-only until the normal training and publication gates pass.
scripts/build_kaggle_proof_training_bundle.py is the deterministic proof
bundle producer for the public Kaggle preview dataset; it emits synthetic,
visible-rationale SFT/preference rows plus held-out splits for the same release
gate.
scripts/build_multiperspective_training_bundle.py is the larger deterministic
case-graph producer. It exposes shared dated synthetic records through single,
handoff, and multi-actor views across perspectives, journey stages, temporal
lenses, evidence states, and jurisdiction topologies. It holds out complete
mechanism families, emits an executable quality audit, and requires a separate
manifest-bound publication approval before the normal release gate can run.
scripts/build_kaggle_interim_collection.py reverifies the approved proof
release and builds two exact-row Kaggle dataset views plus CPU integrity,
training-plan, and four-arm-evaluation notebooks under gitignored reports/.
It does not approve new rows, publish the larger candidate, enable GPU training
by default, or represent visible rationales as hidden chain-of-thought.
Canonical Lane Language¶
Use "six lanes" for the public story, in this exact order:
- Platform safety
- NGO & regulator
- Individual worker / mobile
- Researcher
- Anonymized knowledge sharing
- Developer / integration partner
Use "workflows" only for runnable surfaces such as Chat, Harness Comparison, Bulk File Review, Knowledge Extraction, Search, and Anonymization & Sharing. Do not describe the public story with an outdated lane count.
Kaggle Surface Language¶
Active Kaggle submission surfaces:
kaggle/01-duecare-exploration-workbenchkaggle/02-live-demokaggle/A-00-omni-experiment-workbench
Optional benchmark surfaces:
kaggle/03-universal-llm-benchmarkkaggle/04-kaggle-community-benchmark
Shared dataset surfaces are not additional kernels. In particular,
kaggle/shared-datasets/training-data
is a documentation-only template for a future advanced SFT/preference release.
It must not contain training rows or an active dataset-metadata.json until a
clean manifest-bound release has passed provenance, licensing, privacy,
lineage, held-out, and quality gates.
The published proof dataset
taylorsamarel/duecare-proof-finetuning-data
is separate from that template. It is a preview artifact generated from
scripts/build_kaggle_proof_training_bundle.py, not the full advanced corpus.
Its public SFT and preference views and three completed CPU companion notebooks
are listed in kaggle/_INDEX.md; they are auxiliary proof artifacts and do not
change the three-kernel active submission count.
Archived reference material:
kaggle/_archive/notebooks/03-duecare-video-pitchkaggle/_archive/notebooks/A-01throughA-24kaggle/_archive/notebooks/A-30-cot-gpu-training
Do not add appendix A-* folders at the root of kaggle/ except the active
A-00-omni-experiment-workbench. The only root 04-* folder should be
kaggle/04-kaggle-community-benchmark; other 04-* notebook snapshots belong
under kaggle/_archive/notebooks/.
Commit Checklist¶
Before committing files that affect public review:
- Confirm the file has a clear purpose line or is listed in a purpose map.
- Check for stale lane-count wording.
- Run
python scripts/validate_public_surface.py. - Run a targeted test for the changed surface.
- Keep unrelated generated reports unstaged.