Skills and reusable modules

Agent and skill ecosystem

Codex and Antigravity provide the agentic workspaces. Skills add focused procedures that can be reused by either environment; they do not replace validation evidence or human approval.

Skills invoked in the S6E2 case study

The project installed 166 K-Dense Scientific Agent Skills and one NVIDIA Kaggle skill under .agents/skills/. Installation does not imply use: the audited workflow invoked only the relevant subset.

Skill	How it influenced the work
Codex skill-installer	Installed GitHub skill repositories project-scoped and verified manifests
NVIDIA nvidia-kaggle-skill	Structured Kaggle discovery, kernel push/poll/download, quota inspection, and guarded submission
K-Dense exploratory-data-analysis	Guided aggregate-safe dataset inspection and leakage review
K-Dense scikit-learn	Guided stratified validation, ROC-AUC, and OOF evaluation
K-Dense statistical-analysis	Guided fold-delta and stability interpretation for very small gains
Browser control	Diagnosed the signed-in Kaggle accelerator state when metadata and runtime disagreed
openai-docs	Used for Codex/Headroom configuration checks, not for model selection
imagegen	Produced an earlier banner concept, later replaced by the code-native SVG; it did not affect modeling

The skills supplied procedures and safety boundaries. They did not replace measured experiment evidence.

Competition modules

Frozen fold construction — src/core/cv_folds.py

Creates deterministic stratified fold assignments shared by all model families. Reusing the same fold IDs permits row-for-row OOF comparison.

Candidate evaluator — evaluate_oof_candidate.py

Checks ID/target/fold alignment, rank-transforms predictions, selects a candidate weight on non-held-out folds, and scores the held-out fold. This is the central protection against reporting an in-sample blend optimum.

RealMLP trainer — train_v7_realmlp.py

Supports the representations tested from V7 through V17:

• raw all-categorical;
• hybrid numeric/categorical;
• raw plus bin/digit features;
• optional original-source statistics;
• singleton-only or singleton-plus-pair original groups;
• configurable seed, folds, n_cv, n_ens, epochs, and GPU requirement.

CatBoost trainer — train_v7_catboost_ordered.py

Provides shallow Ordered/Plain CatBoost candidates on the same canonical folds. CPU runners made these long jobs independent of the user's local machine.

Feature engineering — src/core/feature_engineering_v2.py

Creates multiple views of continuous variables, including bins, rounding, and digit/remainder structure. These views were useful because the Playground data was synthetic and retained discrete generator artifacts.

In-loop target encoding — src/core/in_loop_target_encoder.py

For category c, count n_c, category target mean ȳ_c, global prior μ, and smoothing m:

TE(c) = (n_c × ȳ_c + m × μ) / (n_c + m)

Mappings are learned on the active training partition and frozen before validation transformation.

Rank and ridge ensembling — src/core/ridge_ensemble.py

Rank blending was preferred for ROC-AUC because it combines ordering rather than incompatible probability scales. Ridge/logistic stacking remains available but was not the final V17 combiner.

Submission gate — src/core/verify_submission.py

Checks:

• exact row count;
• ID and column alignment;
• absence of NaN/Inf;
• one prediction per test row;
• valid numeric bounds.

Training and submission are intentionally separate operations.

Project-scoped skill provenance

• K-Dense-AI/scientific-agent-skills
• NVIDIA/nvidia-kaggle

K-Dense method citation:

Kassis, T., Agarwal, V., He, Y., Patel, D., & Brueckner, A. M. (2026). Scientific Agent Skills: Library of Procedural Knowledge for Research Agents. arXiv:2609.00065. https://doi.org/10.48550/arXiv.2609.00065