Hello AutoSci team,

First of all, thank you for sharing AutoSci. I am a nutritionist by background, and I am very much an outsider to computer science. My understanding of AI agents, Skills, and related research is still quite shallow, so please forgive me if my wording is not technically precise.

Recently I have been reading AutoSci, some related papers/articles about agentic workflows and Skills, and your wiki-centered Skill design. They have been very inspiring to me. Around the same time, I had also been thinking from my own nutrition work experience about how an AI workflow might first follow the most relevant path, but then still check nearby risks so that important edge cases are not missed. So the thought below simply comes from my day-to-day work experience in nutrition, plus a beginner-level association I made while learning from these materials.

I noticed that AutoSci already has a thoughtful citation discipline / citation-verification reference. I especially like the idea that citations should come from verifiable sources rather than LLM memory, and that uncertain BibTeX entries should be marked as [UNCONFIRMED]. This feels very important.

While reading this, I wondered whether there might be a related but slightly broader pattern for high-risk scientific domains.

In nutrition and medical-related work, after producing an assessment, explanation, or recommendation, we often cannot stop at “the citation exists.” We also need to ask questions such as:

• What level of evidence is this claim based on? A formal guideline, peer-reviewed paper, preprint, official web page, news article, blog post, or just a search hint?
• Is the claim being generalized beyond the population it came from?
• Are there special populations or contraindications that should be checked? For example children, pregnancy, diabetes, kidney disease, drug interactions, etc.
• Are there boundary conditions, counterexamples, or red flags that a first-pass answer might miss?
• Should some claims be softened, flagged for human review, or removed?

From my limited understanding, this feels like a small “missed-case sweep” after the main reasoning path. In human nutrition work, we first identify the likely main problem, but then we still sweep nearby risk areas so we do not miss important contraindications or special cases.

This also made me think of Skill / agent workflows in a very rough way:

1. Main path first — use the most relevant Skill / wiki context for the current task.
2. Nearby risk sweep — after the main draft or plan, check a small set of neighboring risks or evidence boundaries that are easy to miss.
3. Evidence boundary labels — clearly distinguish guideline / peer-reviewed paper / preprint / official page / blog / search hint / unverified claim.
4. Budgeted references — keep the first check lightweight, and only open heavier references when a risk or uncertainty is triggered.
5. Review log — briefly record why a risk check was triggered and what was flagged for human review.

I am not suggesting that AutoSci should become a medical or nutrition system. My thought is much smaller and more general: perhaps AutoSci could have an optional shared reference, or a small review skill, for “evidence boundaries and domain-risk sweep” that complements the existing citation-verification discipline.

One possible minimal form might be:

.claude/skills/shared-references/evidence-boundary-and-risk-sweep.md

It could define a general checklist such as:

1. classify evidence level: peer-reviewed / preprint / guideline / official page / blog / search hint / unverified;
2. identify whether the artifact touches a high-risk domain: medicine, nutrition, law, finance, safety-critical engineering, etc.;
3. run a missed-case sweep after a draft: boundary conditions, special populations, contraindications, counterexamples, and claims that require human review;
4. write a small review log explaining which risk checks were triggered and why.

A slightly larger version might be a small Skill such as:

/evidence-risk-review

It could read a wiki idea / paper plan / draft / review artifact, then write a report under wiki/outputs/ and append to wiki/log.md. But I am not sure whether this fits AutoSci’s current design philosophy.

So this is only a gentle question / proposal:

Would a general “evidence-boundary + risk-sweep” shared reference or small review Skill be aligned with AutoSci? If the idea feels useful, I would be happy to think about a very small PR that follows your bilingual Skill structure and testing rules. If it does not fit the project direction, that is completely understandable too.

Thank you again for the inspiring work.