Skip to content

v2.0.0 — Agent ensemble & risk-proportional leakage detection

Choose a tag to compare

@jiachengwang-punch jiachengwang-punch released this 30 May 02:27
· 4 commits to main since this release

Major release introducing a 4-agent ensemble and reframing data-leakage detection to be risk-proportional.

Added

  • 4-agent ensemble for more comprehensive, rigor-checked analysis (portable role-prompts in references/agents/ as the single source of truth, plus Claude plugin wrappers in agents/):
    • methodology-architect — pre-modeling analysis blueprint; DL-aware (DL is a must-be-justified candidate on tabular data, GBDT usually wins); emits a leakage-risk rating + list.
    • model-diagnostician — adversarial post-evaluation failure hunting; explicit overfitting (train-vs-test) check distinguished from leakage; model-family-agnostic.
    • tool-scout — web-search-backed library/model recommendation (LightGBM is the default, not automatically best); clearly-flagged offline fallback.
    • professor-reviewer — professor-perspective final review with grade + remediation checklist; does not auto-rerun.
  • .claude-plugin/plugin.json so the repository can be used as a Claude Code plugin (agents available natively).
  • Agent-ensemble sections in SKILL.md, METHODOLOGY.md, and README.md.

Changed

  • Reframed data-leakage detection from an absolute, always-mandatory step to risk-proportional screening: the screening question is always asked (near-zero cost), but investigation depth scales with data provenance. Aligned with the existing "match method to constraint" principle.

Full changelog: see CHANGELOG.md