Calma

Independently verify a computational result by RE-EXECUTING it to ground truth and recomputing the headline number from raw outputs - then prove or break the claim. Use to check what an AI agent just produced (a metric, a backtest, a cleaned dataset, a "tests pass"), or as an inline guardrail an agent calls while it works. Recompute-and-diff against the claim + trivial-baseline edge, across domains and languages (quant Sharpe/return/drawdown; classification accuracy/AUC/F1/precision/recall; regression RMSE/MAE/R2; analytics row-count/column-sum/mean). Runs Python/R/Julia/C++/Rust as a black box. Deeper validity checks - leakage re-run, deflated-Sharpe/overfitting, realism deflators, contamination - are named roadmap (M3-M4), not yet delivered. Triggers: "verify", "is this result real", "did the number reproduce", "recompute this metric", "check this backtest/AUC/accuracy", "I processed N rows", "the total is", "before I report this".

mkdir -p ~/.claude/skills/calma
curl -L https://claudskills.com/skills/calma/SKILL.md \
  -o ~/.claude/skills/calma/SKILL.md

@misc{calma-2026,
  author    = {rikhinkavuru},
  title     = {Calma [Claude Code skill]},
  year      = {2026},
  publisher = {ClaudSkills},
  url       = {https://claudskills.com/skills/calma/}
}