Home · ← VS Code | 4/8 | Consensus →

Semantic Scholar

Free, unauthenticated, and scriptable — the literature tool with a script

Overview

Semantic Scholar

The one literature tool this repo actually wraps in a script, because it's free and needs no key.

What it is

What it is

  • Academic search engine and citation graph (semanticscholar.org).
  • Public API, no key needed at classroom volume.
  • Wrapped here by scripts/semantic_scholar_search.py.
Where it fits

Where it fits

D1·1Environment check
D1·2AGENTS.md
D1·3Literature tools
D1·4Referee skill
D1·5Subagents
D2·1Hypothesis
D2·2Design
D2·3Data lab
D2·4Results
D2·5Manuscript
D2·6Governance

The base layer every other literature-verification step in the course builds on.

Access & cost

Access & cost

Free, no account needed for search. That's exactly why this repo ships a script for it and not for Consensus/Elicit/Scite.

In the course

D1·3 — the core exercise

  1. Run python scripts/semantic_scholar_search.py "<query>".
  2. Results save to literature_results.json.
  3. Pick the 3 most relevant results to check further.
In the course

The rule it exists to support

AGENTS.md: never invent citations. A paper only gets cited if it comes from an actual search result — this file is that result.

In the course

D1·5 — literature-scout

Reads literature_results.json (or a Consensus/Elicit/Scite result you pasted in) before citing anything. If it can't verify a claim this way, it says "unverified" — explicitly, not softened into a plausible citation.

In the course

D1·4 — the referee skill's citation check

Check 3 of /referee-review's pipeline cross-checks every citation in a draft against this file. Anything not traceable to it (or another real search) gets flagged as unverified.

In the course

D2·5 — manuscript citations

Every citation that makes it into the final draft should trace back to this file, not to the model's memory of what a paper probably says.

Limits

Limits

  • Mostly abstracts and open metadata, not full text.
  • Doesn't tell you whether a paper's findings have been contradicted since publication — that's Scite's job.
  • Doesn't synthesize across papers — that's Consensus/Elicit's job.
How it fits together

How it composes with the other three

Semantic Scholar finds candidates → Scite checks whether they've held up → Consensus/Elicit's synthesis gets checked against the actual abstract. Four tools, one discipline: verify before you cite.

Recap

Recap

See D1·3's decision graph on the home page for the full verify-before-citing flow.