Kaggloop Hypothesize
Stage 2 of the kaggloop win-loop and its highest-leverage stage — where the competition is won or lost. Begin each round with a mandatory re-recon — the iron rule first (sync + read the top-5 best-Public-Score notebooks via kloop.notebooks, byte-deduped; enforced at stage close), then the leaderboard, discussions, and fresh papers, driven by the gap and prior iterations; log it to the cumulative…
Kaggloop Survey
Stage 1 of the kaggloop win-loop — deep-dive the chosen competition into one dossier (data, exact metric, a leakage-safe CV scheme, rules, the top-5 best-Public-Score notebooks synced locally via kloop.notebooks (the iron rule, enforced) with the best one designated the baseline to adapt and beat, key discussions, relevant papers via the science MCP) AND set the target score the loop will chase (…
Kaggloop Publish
Upgrade the public Kaggle notebook to a user-specified bundle with a matching Public-Score badge. Use whenever the user says to publish / upgrade / "notebook-ify" a bundle or a new best score, or asks to put a specific score on the public notebook. Handles slug resolution, content-match, badge, and vote preservation automatically.
Kaggloop Submit
Stage 4 of the kaggloop win-loop — finalize the round by ensembling the kept models, passing the strict data-leakage gate (enforced before any submission), submitting to Kaggle via the API — the standard ensemble AND the round's challenge-track submission (the dual-submission mandate, enforced at stage close) — recording the public leaderboard scores, comparing the actual score to the target, run…
Kaggloop Scout
Stage 0 of the kaggloop win-loop — turn a competition the user is interested in (a Kaggle URL or slug) into a project plus a human-readable TLDR card for a go/no-go decision; or, in discovery mode, shortlist several candidates. Use at the very start, when the user pastes a competition URL, or asks "what should we compete in?". This is the one mandatory human-in-the-loop gate; do not auto-select.
Kaggloop
Orchestrate an end-to-end Kaggle competition project — scout a competition (human picks one from TLDR cards) → survey it + the literature and set a target score → form critical-to-win hypotheses → run experiments on Colab → pass a strict data-leakage gate → ensemble & submit, then loop while the score is short of the target. Use when the user wants to "run kaggloop", autonomously enter/win a Kagg…
Kaggloop Experiment
Stage 3 of the kaggloop win-loop — implement and verify the top-ranked hypotheses by writing pipeline code (starting from the synced best-Public-Score public notebook as the baseline — the iron rule; never scratch-written code below the public floor), running it on Google Colab (GPU) via the kloop.colab bridge, scoring with the dossier CV, and running the data-leakage gate on each result before k…