hs:bakeoff

What problem does it solve? When reasoning cannot separate two or more viable technical directions, teams often pick a winner by argument alone. This Skill replaces opinion with evidence by building a cheap throwaway probe per candidate, running every probe on the same input set with the same mechanical metric, and letting the numbers decide — or honestly declaring a tie when the gap falls inside the measured noise band. ## Core Features & Use Cases - Empirical candidate comparison: Fans out 2-4 candidates into isolated git worktrees, measures each with an identical shell-command metric, and records every trial in an append-only JSONL ledger. - Honest verdicts with noise handling: The rank script refuses to crown a winner when the gap is within the noise band (relative band or observed spread) or when trial counts are insufficient, handing the decision to a human instead. - Gate-first preflight: Validates candidate count, metric safety, and budget ceilings before any probe is built, and falls back to cheaper reasoning when a bake-off is not justified. - Use Case: Deciding between redis, in-memory, and sqlite caching — build a minimal adapter probe for each, measure latency on the same input set across multiple trials, and get a verdict showing the full scoreboard including losers and spread. ## Quick Start Ask the AI to bake off your candidate approaches, for example: run a bake-off comparing redis, in-memory, and sqlite on query latency with medium noise and a 180-second budget per probe.