Erlemar

Use this skill for any task in the cayley IHES Picture Cube project — training, solving, evaluation, ensembling, submission. Covers the full ML+search pipeline.

Strat-5 acceptance gate for a megaminx V model. Runs 51-pid stratified eval at beam 65k, compares to m05 baseline (50/51 / mean 89.4), reports pass/fail.

Run the top-N long-pid sym-ensemble rescue + merge cycle on the current best megaminx submission.

One-shot status across GCP solves, Kaggle kernels, and local runs for the megaminx competition.

Verify a megaminx submission CSV, submit to Kaggle, and update HANDOFF / SUBMISSION_JOURNEY / to_do_shortlist with the new score.

Pull the latest cayleypy-tpu-beam-smoke kernel output, aggregate per-rank JSONs, merge with current best, verify, and prep for submission.

Combine candidate submissions, post-process, verify, and submit to Kaggle.

Verify an already-built submission CSV then submit to Kaggle.

Combine all model OOF predictions into the best validated blend/stack, apply the frozen recipe to test predictions, and write submission.csv. Use after the model agents finish, or any time the OOF pool changes.

Create or validate the shared cross-validation split (artifacts/folds.parquet) that every model must use, and check that it tracks the leaderboard. Use before any model trains, or whenever CV and LB diverge.

Scaffold a new tabular-multiclass competition in this kit — download/place data, detect id/target/classes/metric, fill src/config.yaml, pick the fold scheme, and freeze the shared CV split. Use once at the very start, before EDA.

Validate a submission.csv against sample_submission before spending a scarce daily submission slot — column names/order, id coverage, prob ranges/sums, NaNs, and canonical class order. Use right before every leaderboard submission.

Generate Excalidraw `.excalidraw.md` source files for DSWoK vault diagrams (charts, pipelines, conceptual sketches). Use ONLY when the user explicitly asks for an Excalidraw diagram or when the diagram-tooling order in `.claude/writing_style_guide.md` selects Excalidraw (non-flowchart visuals where Mermaid is the wrong tool).

Extract patterns and learnings from the current session. Use when asked to learn from this session, capture session learnings, or review what was accomplished. Outputs structured findings for use with /llm-rules-updater.

Detect and roast AI code slop - redundant, unreadable, or unnecessarily complex code patterns. Focuses on stupid comments, sloppy tests, over-abstraction, and repetitive code that makes codebases painful to maintain.