Cross-competition evolution playbook

The Arena evolves through evidence accumulated across competitions. Detailed version histories belong to their case-study pages; this page contains the rules carried into the next challenge.

Establish truth

• Audit data provenance, target, metric, and leakage risks.
• Reconcile documentation with actual OOF and leaderboard artifacts.
• Freeze folds, IDs, target order, and artifact schemas.
• Create a cheap reproducible baseline before expensive search.

Search for complementary signal

• Change one major factor per experiment.
• Explore genuinely different representations and model families.
• Measure fold-level deltas and correlations against the current ensemble.
• Use cross-fitted or otherwise held-out blend selection.
• Prefer a small stable gain to a complex in-sample optimum.

Operate remote compute safely

• Match workloads to Kaggle CPU, GPU, or TPU.
• Fail fast when the requested accelerator is unavailable.
• Save fold checkpoints and a run manifest.
• Poll and download artifacts without coupling training to a local laptop.
• Never let a training kernel submit automatically.

External evidence and reporting

• Run an integrity gate before upload and submit only after human approval.
• Record score type, submission ID, date, and competition state.
• Keep leaderboard regressions in the experiment ledger.
• Publish one dedicated report per competition and feed reusable lessons back here.

First case study

Playground Series S6E2 showed that measurement repair, representation diversity, generator-aware features, and cross-fitted rank blending mattered more than blindly increasing model size. Several plausible variants were rejected; V17 matched the displayed winning Private score. The full progression, prompting approach, skills, compute history, token snapshot, and limitations are in Playground-Series-S6E2.