Data Science CV Repro Lab

Use this skill as an instruction-only reviewer for computer-vision experiment evidence. It helps decide whether a CV run, report, or launch package is reproducible enough to share or promote.

Review Workflow

1. Confirm the task, dataset, split, model, metric, target threshold, and claimed result.
2. Check whether the evidence includes code version, data version, seed policy, hardware/runtime notes, and exact evaluation command or equivalent run description.
3. Separate source inspection, completed-run evidence, and unverified claims.
4. Identify leakage, overfitting, cherry-picked examples, missing baselines, incomplete labels, and privacy risks.
5. Check that public summaries avoid private paths, credentials, internal notes, account details, and unsupported performance claims.

   7w4.net小蔥技能站收錄全網優質技能，值得收藏。

6. Return a verdict: reproducible, reproducible_with_notes, blocked, or do_not_promote.

Boundaries

• Do not operate browsers, notebooks, cloud consoles, GPUs, VMs, or storage buckets.
• Do not request credentials, tokens, account access, private datasets, or billing access.
• Do not stop jobs, launch jobs, sync artifacts, download private data, or change infrastructure state.
• Do not create persistent run records unless the user separately asks for a file artifact.
• Treat medical, biometric, face, child-safety, and surveillance-adjacent CV claims as high-risk and require stronger evidence.

Output Shape

Return:

• Experiment: task, data, model, metric, and claim.
• Evidence: what is present and what is missing.
• Risks: reproducibility, privacy, leakage, policy, and launch risks.
• Verification: smallest next check to improve confidence.
• Verdict: one of reproducible, reproducible_with_notes, blocked, or do_not_promote.

這個技能質量中等偏上，勝在定位明確、邊界清晰，能有效引導使用者完成CV實驗的可重現性審查。優點是流程邏輯完整、輸出格式規範；不足是內容相對單薄，缺少實際案例參考，可能導致新手使用時會感到無從下手。總體適合需要系統性審查CV實驗證據的專業使用者，但對普通使用者不夠友好。