AI 工程从零到一（中文）

Given an autonomous research output (hypothesis, code, experiments, figures, paper draft) produced by an AI-Scientist-v2-style loop, produce a two-gate review: sandbox audit (does anything leave?) plus research audit (is the work sound?).

The two gates map directly onto the audits below: Sandbox gate = item 1; Research gate = items 2 (Experiment audit) + 3 (Polish audit). Items 4–5 govern what happens after both gates pass.

Produce:

1. Sandbox gate. Before any artifact leaves the sandbox:
   • List every network call the loop made and its target. Flag any that were not pre-approved.
   • Inventory every file the loop wrote outside its working directory.
   • Confirm Docker / seccomp / gVisor containment held for the full run.
   • Confirm no subprocesses escaped the sandbox's supervision. If any check fails, block export; raise to a human.
2. Experiment audit. Read the experiment code, not the paper:
   • Verify every claimed experiment actually ran and its reported numbers are reproducible.
   • Check that failed experiments were reported as failures, not re-framed as negative results after-the-fact.
   • Check that the "novelty" label on the idea holds up against a literature search by a human domain expert.
3. Polish audit. Read the figures:
   • Ensure every figure's data came from a logged experiment run, not from polish-stage rewriting.
   • Confirm axes, scales, and annotations match the underlying data.
   • Flag any figure whose caption claims more than the data supports.
4. Disclosure plan. If the artifact is intended for external distribution:
   • Disclose that the artifact is agent-authored.
   • Disclose the tools used (model family, loop version).
   • Disclose the human reviewer who checked it and what they checked.
5. Negative-release decision. If the artifact fails any audit step, the default is do not release. Overriding this default requires a named human owner.

Hard rejects:

• Any submission that skips either gate.
• Any artifact where the loop's execution logs are missing or incomplete.
• Any figure that cannot be traced to a specific experiment run.
• Any novelty claim that a domain expert has not verified.

Refusal rules:

• If the run lacks Docker or equivalent isolation, refuse and require re-run in an isolated sandbox.
• If the user cannot produce execution logs for the experiment stage, refuse — the paper is unreviewable.
• If the proposed distribution channel is a peer-reviewed venue and the user proposes not to disclose agent authorship, refuse and require disclosure.

Output format:

Return a two-gate report:

• Sandbox gate verdict (PASS / BLOCK, with rationale)
• Research gate verdict (covers Experiment audit (2) and Polish audit (3)) (PASS / BLOCK / REQUIRES_EXPERT, with per-check notes)
• Disclosure plan (venue, text, human reviewer name)
• Release decision (release / hold / reject)
• Next action (who does what by when)