command-result-promotelisted
Install: claude install-skill a-green-hand-jack/ml-project-repo-agent-native-template
# command-result-promote
Codex does not load project `.claude/commands/*.md` files as custom slash commands.
Use this skill when you would have used `/result-promote` in Claude Code.
Canonical source: `.claude/commands/result-promote.md`. Do not edit this adapter by hand; edit the
Claude command and run `python scripts/sync-codex-adapters.py`.
评估把 $ARGUMENTS 的结果 promote 进 `lab/research/evidence.yaml` / `lab/artifacts/result-index.yaml` / 论文。
先检查证据门槛(见 `.agent/artifact-policy.md`):run 可定位、config 可复现、metric 来源清楚、
与 baseline 比较清楚、caveat 写明。
promote 前跑 `python scripts/check-provenance-chain.py`:run 必须已闭环(experiment-ledger
`status: done` + `run_summary`),result-index 条目 checksum(统一 sha256)与引用链必须通过;
无法校验的 checksum 需填固定枚举 reason + 非占位人工理由(见 `.agent/artifact-policy.md`)。
promote 是 human gate(见 `.agent/human-gates.md`):用 tier 4 fresh verifier 复核后,
附 run id / config / commit / checkpoint / data split / metric source,再请求人工批准。
不要在有旧训练日志的当前上下文里直接下 claim。
证据门槛与 evidence 写入的完整流程见 `experiment-workflow` skill;本 command 只是它的 promote 入口。