Skip to content

Readiness Dashboard

Current as of 2026-05-17. This replaces the historical 2026-05-02 appendix-ladder dashboard; the active submission scope is now the active Kaggle Gemma 4 path.

Active Scope

Surface Status What It Proves
kaggle/01-duecare-exploration-workbench/ Active Interactive harness comparison, chat, extraction, search controls, traces, and knowledge-pack flows.
kaggle/02-live-demo/ Active Focused live demo and video narrative path.
kaggle/A-00-omni-experiment-workbench/ Active Quantitative proof path: baseline vs harnessed arms, synthetic data, fine-tune, judging, and report artifacts.

Archived A-series notebooks, task-notebook snapshots, and old generated mirrors are not the active competition path. Root kaggle/ should not contain appendix A-* folders, and the only root 04-* folder should be 04-kaggle-community-benchmark. See current_kaggle_notebook_state.md and kaggle/_INDEX.md.

Current Green Checks

Area Current State
Harness contract Documented in harness_ecosystem.md, harness_pattern.md, and harness_standard_contract.md.
Model loading Standardized through Gemma4Runtime.load() for inference, with active A-00 training as the only direct FastModel exception.
Archived A-00 default harness chat_no_online: Persona + GREP + RAG/context + deterministic tools, with internet/import off.
Archived A-00 judging Combined rule + LLM judging with local Gemma by default and optional external judge adapters.
Archived A-00 exports HTML, Markdown, JSON, CSV, charts, activity/evidence bundles, and report manifest under /kaggle/working.
Test baseline Focused contract gates are listed in FOR_PEER_REVIEW.md.

Remaining Human Actions

  1. Run the two active Kaggle kernels on the intended GPU/runtime shape.
  2. Optional only: produce an A-00 evidence run with checkpoint/resume enabled if new proof artifacts are needed.
  3. Download /kaggle/working artifacts before Kaggle shutdown.
  4. Capture final demo screenshots/video clips from Kernel 01 and Kernel 02.
  5. Submit with links pointing to the current active Kaggle path.

The manual checklist lives in USER_TODO.md.

Archived A-00 Evidence Run Targets

For a fast proof run, use 4 prompts and training disabled or a very short LoRA smoke path. For a writeup-quality run, use the highest prompt count that fits inside the remaining Kaggle wall-clock budget, keep checkpoints enabled, and prefer a larger Gemma/frontier judge only if credentials and runtime allow.

Required exported evidence:

  • Full activity log with all step details.
  • Prompt/response JSONL for every arm.
  • Synthetic SFT rows when generation is enabled.
  • Training config, checkpoints, adapter path, and resume metadata when training is enabled.
  • Final rule + LLM judging rows.
  • HTML/Markdown/JSON report plus charts and manifest.

Current Risks

Risk Mitigation
Kaggle runtime hits the time limit Keep checkpoint/resume enabled and save artifacts after each major phase.
A-00 judging takes too long Use local small Gemma for proof runs; reserve larger/frontier judges for final scoring or post-run regrade.
Old docs confuse reviewers Active entry docs now point to the current Kaggle scope; legacy roadmap docs live under docs/_archive/.
Harness parity drifts Contract tests pin Kernel 01 and active A-00 parity for runtime loading, default harness layers, and shared GREP/RAG/tool usage.

Start Here