thalarch-autoresearchlisted

# Thalarch Autoresearch Autoresearch is a **bounded engineering experiment protocol**, not permission for an agent to edit indefinitely until a number happens to look better. Use it when the task has a meaningful evaluator and several plausible changes can be tested against that evaluator. The loop must remain subordinate to the user's scope, repository rules, correctness, safety, and external-action boundary. ## 1. Activation gate Use this skill when at least one is true: - the user asks to optimize a measurable property; - several plausible implementations can be compared objectively; - a performance/build/startup/latency problem has a reproducible benchmark; - a difficult bug has repeated falsifiable hypotheses and a stable reproduction/test; - an agent, prompt, skill, heuristic, or workflow is being tuned against a frozen evaluation set; - an existing benchmark suite can discriminate candidate changes. Do **not** activate merely because a task is difficult. Prefer ordinary `thalarch-debug`, `thalarch-performance`, `thalarch-test`, or `thalarch-code-craft` when one well-supported change is enough. Do not run an optimization loop when the evaluator is subjective, easily gamed, unstable, or materially weaker than the real requirement. Improve the evaluator first or keep the result `UNVERIFIED`. ## 2. Research contract Before the first candidate, freeze a compact contract: - **objective** — what real outcome should improve; - **primary metric** — exact name, unit, d