research-proposal

Provide your draft proposal and let the loop run ScholarEval for literature-grounded critique, a Judge for a 0-100 grade against a fixed rubric, and a Reviser that applies ranked fixes. The loop repeats until the grade passes your threshold or the iteration budget is exhausted.

The rubric scores three axes from 0 to 5: soundness, contribution, and evidence quality, then aggregates them into a 0-100 grade with frozen weights (0.45, 0.35, 0.20 by default). Hard gates force a fail if soundness is zero, evidence quality is one or below, or a fatal ungrounded contradiction exists.

No. The skill requires an existing proposal containing a problem statement, proposed methodology, and planned experiments. It also does not run a standalone literature survey; retrieval exists only to ground the evaluation and revisions.

No key is required for the core Semantic Scholar and arXiv search, snippet, cite, and fulltext commands. A free S2_API_KEY makes snippet and citation-graph retrieval more reliable, and the loop degrades to WebSearch and WebFetch if the literature tools are unavailable.

Every evidence entry must contain a verbatim snippet from a real retrieval performed that iteration, and the Judge applies an evidence gate that discounts ungrounded points before scoring. Claims without real citations are marked ungrounded rather than given invented references.

The loop stops when the verdict passes the grade threshold, when the iteration budget is reached, or when the grade has not improved for the configured patience count. On budget or plateau stops it reports the best-grade iteration, not necessarily the last.