Skip to content

Final Submission Execution Plan

Current as of 2026-07-26. This document replaces the old 2026-05-02 two-week appendix-ladder plan. The remaining work is no longer broad notebook publishing; it is validating the active kernels and preserving optional evidence from the active A-00 pipeline.

Scope

Run and verify:

  • kaggle/01-duecare-exploration-workbench/
  • kaggle/02-live-demo/
  • kaggle/A-00-omni-experiment-workbench/

Everything under kaggle/_archive/notebooks/ is historical unless explicitly revived later, including legacy notebook wrappers and task-notebook snapshots. Root kaggle/ should not contain appendix A-* folders other than active A-00, and the only root 04-* folder should be 04-kaggle-community-benchmark.

Priority Order

  1. Run the model-free core publication gate and freeze the intended release commit, versions, citation metadata, and claim set.
  2. Confirm the official Kaggle submission requirements and final deadline.
  3. Decide whether to publish the existing bounded evidence now or first clear the separate 75-row corridor-diversity training blocker.
  4. Run Kernel 01 and capture the default harness comparison path only if a new recording is needed.
  5. Run Kernel 02 and capture the live demo/video path only if a new recording is needed.
  6. Optional only: run active A-00 with a small proof configuration to confirm exports and report links if new proof artifacts are needed.
  7. Optional only: run active A-00 again with the best available prompt count/training settings that fit the remaining wall-clock budget.
  8. Download /kaggle/working outputs after every meaningful phase.
  9. Submit with the current active Kaggle story, not the archived A-series story.

Active A-00 Runtime Strategy

Use the preconfigured pipeline for the main proof. Keep options narrow:

  • Small Gemma 4 model for generation/fine-tuning smoke runs.
  • Offline harness default: Persona + GREP + RAG/context + deterministic tools.
  • Combined rule + LLM judging at the end.
  • Checkpoint/resume enabled for training.
  • Larger Gemma/frontier/Ollama judge only if credentials and runtime allow.

If the run approaches Kaggle's time limit, stop at a saved checkpoint or after a completed response/judging phase, download artifacts, and resume in a later session.

Evidence To Preserve

  • Activity log export.
  • Run manifest and pipeline request.
  • Prompt/response artifacts for all arms.
  • Harness traces for harnessed arms.
  • Synthetic SFT rows.
  • Training script, config, logs, checkpoint metadata, and final adapter path.
  • Judge model configuration and per-response combined scores.
  • HTML, Markdown, JSON, CSV, chart, and ZIP report artifacts.

Do Not Spend Time On

  • Rebuilding archived notebook-era surfaces.
  • Reintroducing duplicate model-selection UI.
  • Refactoring the canonical GREP/RAG corpus during final proof runs.
  • Spending Ollama quota before an offline plan, frozen prompt/model/rubric scope, finite call allowance, and checkpoint-reuse decision exist.
  • Adding broad new docs that compete with FOR_PEER_REVIEW.md, USER_TODO.md, and this execution plan.