[Visual notebook] How 203 experiments reached public #16: the complete idea map
Update: the notebook is now a full visual field guide, not a five-chart summary.
An earlier estimator reached public rank #16 under WhestBench 0.13.0 / FlopScope 0.9.1. The evaluator later changed. On the final Phase 1 stack, the frozen method replayed at adjusted score 1.4854216e-7; exact prediction-preserving compute reductions reached 1.4391444e-7 — 3.115426% lower on the same stack — while the growing public board ended near #40. The rank is historical; fresh private re-evaluation decides prizes.
The expanded notebook now includes:
- a 30-second diagram of the estimation problem;
- the official accuracy/compute frontier with evaluator epochs separated;
- a complete map of all 20 structured early branches;
- a visual census and searchable table for all 56 late investigations;
- a “base idea → one change → reason → decision” transition table;
- oracle-capacity vs lawful-deployment and evidence-stage explainers;
- the exact prediction-preserving compiler staircase; and
- a practical FAQ covering ranks, raw vs adjusted score, Kerdock/MUB geometry, negative results, private reruns, and reuse.
The campaign has 203 numbered investigations. The public-safe atlas exposes 20 early branches in grouped form and 56 late rows individually; the notebook says this explicitly instead of inventing 203 point-level records.
Run the executed notebook on Kaggle: How 203 Experiments Reached Public #16
The notebook executes top-to-bottom without challenge data or network access. It contains no targets, per-network identities, private seeds, prize-estimator source, target reconstruction, or accounting-bypass details. ChatGPT/Codex and Claude were used extensively. The decorative hero background was generated with OpenAI image generation; technical diagrams and numerical claims are programmatic and reconciled against frozen receipts or official grader readbacks.
Notebook code is MIT. Narrative and curated data are CC BY 4.0. Technical corrections are welcome.