Kg Ensemble

---
name: kg-ensemble
description: Build the best honest ensemble from ledger experiments (diversity analysis, hill climbing / weights / rank / stacking with nested CV, post-processing re-tune) and write a validated blended submission.
argument-hint: "[experiment ids...] [--method hill|weights|rank|stack|residual] [--prune N]"
allowed-tools:
  - Read
  - Glob
  - Grep
  - Agent
  - 'Bash(PYTHONPATH="${CLAUDE_PLUGIN_ROOT}" python -m kgkit blend *)'
  - 'Bash(python -m kgkit blend *)'
  - 'PowerShell(python -m kgkit blend *)'
  - 'Bash(PYTHONPATH="${CLAUDE_PLUGIN_ROOT}" python -m kgkit ledger *)'
  - 'Bash(python -m kgkit ledger *)'
  - 'PowerShell(python -m kgkit ledger *)'
---

# /kg-ensemble

Run kgkit with the Bash tool, as `PYTHONPATH="${CLAUDE_PLUGIN_ROOT}" python -m kgkit <command>`: this command
pre-approves that form for the kgkit calls it makes. Submitting, pushing or publishing still asks (so does a
PowerShell `$env:` assignment).

Arguments: `$ARGUMENTS` (optional explicit experiment ids; otherwise choose from the ledger).

Delegate to the `ensemble-architect` agent (or follow its procedure directly for small ledgers):
candidate audit (same folds/row order/scale) → correlation & diversity → `python -m kgkit blend
--exp ... --truth data/train.csv:<target> --folds data/folds.csv:fold --method hill --sample
data/sample_submission.csv` (+ alternative methods) → compare in-sample vs honest scores → for large
libraries `--prune 40` then `--method stack --levels 1|2`, or `--method residual --base <id>` → accept the
winner only if `kgkit ledger compare` says it beats the simpler blend → re-tune thresholds/rounding on the
blended OOF → validate the submission.

Blends that include public submission files (no OOF) cannot be scored this way: log them LB-only
(`kgkit ledger add <name> lb-only --sub <file>`, then `kgkit ledger lb`) and follow `ensembling` §6.

Report members, weights, single vs blended (honest) CV, and whether it deserves a submission slot.