Skip to content

Current Kaggle Submission State

This file is the short operational pointer for the current Kaggle scope. Older generated notebook mirrors still exist for provenance and tooling, but they are not the active submission path.

Active Judge-Facing Kernels

The active competition scope is exactly three script kernels, also listed in kaggle/_INDEX.md. The first two are the primary recording path; A-00 is the quantitative proof and training/evaluation path.

Folder Role Live Kaggle status checked 2026-07-28
kaggle/01-duecare-exploration-workbench/ Broad workbench: chat, harness comparison, bulk review, knowledge extraction, search, sharing, traces, and activity logs. COMPLETE
kaggle/02-live-demo/ Focused live demo for judges and video capture. CANCEL_ACKNOWLEDGED
kaggle/A-00-omni-experiment-workbench/ Quantitative proof path: baseline, harnessed, synthetic-data, fine-tuning, judging, checkpoints, and report artifacts. CANCEL_ACKNOWLEDGED

The status values above came from the authenticated Kaggle command line on 2026-07-28. Cancellation is terminal but is not successful completion. A-00's public page still attaches taylorsamarel/duecare-proof-finetuning-data, but its canceled run and unreviewed output boundary cannot be cited as a completed training/evaluation proof. No production Gemma adapter, merged weights, or full advanced training corpus is attached.

Auxiliary Interim Training Proofs

The active judge-facing count remains three. Separately, two exact-row dataset views and three CPU companion notebooks are public:

The training starter's completed public run is a CPU validation and plan export, not a GPU fine-tune. It wrote training-plan.json; the evaluation notebook wrote evaluation-plan.json and evaluation-prompts.jsonl. No adapter weights or claimed model lift are attached.

Each active folder uses kernel.py as source of truth. The existing A-00 Kaggle slug is notebook-typed, so publication packages the script source into a temporary .ipynb wrapper without changing the repository source of truth.

Optional Evaluation Kernels

kaggle/03-universal-llm-benchmark/ is a separate endpoint-comparison kernel. It can test OpenAI-compatible, Anthropic Messages, or raw JSON APIs against DueCare prompts and rubric cues, with Claude Opus judging when configured. It is useful for model comparison work, but it is not part of the primary recording path above.

kaggle/04-kaggle-community-benchmark/ is a separate Kaggle-native benchmark kernel. It defines DueCare rows as kaggle_benchmarks tasks so model calls can use Kaggle's model proxy/quota and publish Community Benchmark task/run artifacts. It is also optional and not part of the primary recording path.

Live review on 2026-07-28 found the Community Benchmark run COMPLETE. The configured Universal LLM Benchmark slug did not resolve publicly and could not be verified; keep it source-only until a public URL returns successfully. The dated closeout decision declined all optional reruns because no named evidence gap justified quota use.

Archived Or Reference-Only Material

  • kaggle/_archive/notebooks/ contains the former 03, A-01 through A-24 notebook-era surfaces, task-notebook snapshots, and provenance copies.
  • Root kaggle/ must not contain appendix A-* folders other than the active A-00-omni-experiment-workbench, and the only root 04-* folder should be 04-kaggle-community-benchmark.
  • kaggle/kernels/ contains generated/research notebook mirror material used by older tooling and tests. It is not the current judge path.
  • kaggle/shared-datasets/training-data/ is a documentation-only template for a future manifest-bound SFT/preference release. It intentionally contains no active dataset-metadata.json and no data rows.
  • _archive/kaggle-notebook-previews-2026-05-11/ contains historical notebook wrappers and metadata snapshots.

Current Review Anchors