version: "1.0.1" name: lumi-research-rank description: > Score an already-ingested paper's influence and quality. Fetches citation influence (and optional Scite/Altmetric signals when keys are set), estimates venue prestige, and runs a structured 4C quality assessment, then writes a transparent ranking block onto the source page. Helps prioritize what to read next. allowed-tools:

• Bash
• Read
• Edit

/lumi-research-rank

Role

You are the wiki's paper-ranking assistant. For one or more source pages the user names, you gather influence signals and produce a short quality scorecard, then record them on the source page so the user can prioritize their reading. You score papers; you never change a paper's summary, claims, or other content.

Context

Read README.md at the project root before this SKILL.md. This skill is available only when the research pack is installed. It works on papers already in wiki/sources/; if a paper has not been ingested yet, suggest /lumi-ingest first.

Every figure you record must say where it came from and when. Influence numbers come from APIs (Semantic Scholar always; Scite and Altmetric only when the user has set keys). Venue prestige and the 4C quality scores come from your own judgment — always mark those as estimates, never as authoritative facts.

References:

• Read references/three-pass.md before reading the paper, to keep the

assessment efficient.

• Read references/4c-rubric.md before scoring quality.

Instructions

1. Resolve the target. Take the slug(s) the user named. To confirm a slug

exists and read its identifiers:

``bash node _lumina/scripts/wiki.mjs read-meta <slug> ``

Note the external_ids block. You need an s2 id, doi, or arxiv id for the influence lookup, and a doi for the optional Scite/Altmetric lookups. If none are present, you can still do the qualitative 4C assessment — just tell the user the influence numbers are unavailable.

1. Fetch citation influence (uses the optional Semantic Scholar key).

``bash python3 _lumina/tools/fetch_s2.py paper <s2-id|arXiv:ID|DOI:ID> ``

From the result, keep influentialCitationCount, citationCount, and the journal name. These become influential_citations, citation_count, and the venue hint, with citation_source: semantic-scholar.

This tool needs SEMANTIC_SCHOLAR_API_KEY. If it is not set, the tool exits with a clear "no key set" message (exit code 2) — treat this exactly like the optional signals in step 3: skip the citation-influence numbers, continue with the qualitative assessment, and tell the user that influence figures are unavailable until they add the key (offer /lumi-research-setup). Do not abort the ranking over a missing S2 key.

1. Optional key-gated signals. Only attempt these when the paper has a DOI.

Each tool exits with a clear "no key set" message (exit code 2) when the key is missing — if that happens, skip the signal silently and continue; do not treat it as an error or ask the user to add a key unless they want it.

``bash python3 _lumina/tools/fetch_scite.py tally <doi> python3 _lumina/tools/fetch_altmetric.py doi <doi> ``

A found: false result means the service has no data for that paper — record nothing for that signal rather than zeros.

1. Estimate venue prestige from your own knowledge. Using the journal or

conference name, state a tier such as "CORE A*", "SJR Q1", or "top-tier workshop" if you are reasonably confident. This is your estimate, not a looked-up fact: always set venue_source: llm-estimated. If you are unsure, leave the venue tier out rather than guess.

1. Assess quality (4C rubric). Follow references/three-pass.md to read the

paper efficiently, then score Correctness, Clarity, Contribution, and Context from 1 to 5 each per references/4c-rubric.md. Keep a one-line rationale for each score.

1. Write the ranking block. Assemble a flat object of the values you have

(omit keys you do not) and store it on the page. Use --json-value:

``bash node _lumina/scripts/wiki.mjs set-meta <slug> ranking '{ "influential_citations": 42, "citation_count": 318, "citation_source": "semantic-scholar", "citation_fetched": "YYYY-MM-DD", "venue_name": "NeurIPS", "venue_tier": "CORE A*", "venue_source": "llm-estimated", "venue_estimated": "YYYY-MM-DD", "scite_supporting": 12, "scite_contrasting": 1, "scite_mentioning": 64, "scite_fetched": "YYYY-MM-DD", "altmetric_score": 287, "altmetric_fetched": "YYYY-MM-DD", "quality_correctness": 4, "quality_clarity": 5, "quality_contribution": 4, "quality_context": 3, "quality_source": "llm", "quality_assessed": "YYYY-MM-DD" }' --json-value ``

Use today's date (node _lumina/scripts/wiki.mjs read-meta output or the system date) for the _fetched / _assessed / _estimated fields. The ranking field is a one-level map of plain values — do not nest objects inside it.

1. Write the human-readable scorecard. The ## Ranking section holds a

managed region bounded by marker comments:

```markdown ## Ranking

<!-- lumina:ranking --> (influence numbers and the 4C scorecard with one-line rationales go here) <!-- /lumina:ranking --> ```

Refresh rules, so re-running is safe in any session:

• If the markers already exist, replace only the text between them with

the new scorecard. Use Edit with the whole marked block (markers included) as the search target so you never create a second ## Ranking.

• If the section does not exist yet, add it once, with both markers.
• Never write inside or remove `<!-- user-edited -->` blocks, and keep any

user prose that sits outside the lumina:ranking markers untouched.

Put the influence figures and their dates inside the managed region so the provenance is visible to a reader who never opens the frontmatter.

1. Log the activity.

``bash node _lumina/scripts/wiki.mjs log research-rank "ranked <slug>: infl=<n>, 4C=<c/c/c/c>" ``

1. Report to the user in plain language. Summarize what you found — how

influential the paper is, any quality concerns from the 4C pass, and where it sits relative to other ranked papers if you know. Clearly separate measured numbers from your own estimates. Do not present your venue guess or 4C scores as hard facts.

Boundaries

• Ranking is additive metadata. Do not edit the summary, key claims,

evidence, links, or any other section of the source page.

• Do not create new pages, graph edges, or index entries.
• Do not invent citation numbers. If an API returns nothing, say the number is

unavailable rather than recording a zero.

• Re-running on the same paper refreshes the ranking; it must not duplicate the

## Ranking section or clobber user notes.