paper-verify

한국어 가이드 → README.ko.md

Fact-check the citations in any document. paper-verify extracts every reference (URL / DOI / PMC / PMID / arXiv) from a Markdown or text file, fetches each source, asks one or more LLMs whether the cited claim is actually supported by the source, scores each citation on a transparent 100-point rubric, and writes a Markdown report that flags fabricated, misquoted, or dead-link citations.

Built for researchers, grad students, lecturers, and bloggers who need to trust their own footnotes — and to catch AI-hallucinated citations before they ship.

It is also a review triage tool: use it to reduce a long bibliography or blog source list into the small set a human should actually inspect. The JSON surface exposes tier, consensus, effective_verdict, judge disagreement, source, landing_status, and soft_404_suspect, so agents can build a shortlist such as Must Review, Review If Important, and Probably Safe instead of asking a human to read every cited source.

No framework dependency

The core has zero required third-party dependencies (Python stdlib only) and no dependency on any agent-orchestration framework. Parallel fetching uses concurrent.futures.ThreadPoolExecutor. LLM providers (Anthropic / OpenAI / Gemini) are optional extras, and a dependency-free keyword judge lets the tool run end-to-end with no API keys at all.

What it does

1. Extract — regex-match citations and capture ~100 chars of surrounding context (the claim being made), with line numbers; deduped by (type, ref).
2. Fetch — resolve each reference and fetch its source through an explicit fallback chain (see below). For academic identifiers (DOI / arXiv / PMID / PMC) it first queries free official metadata APIs (Crossref / arXiv / NCBI) — bypassing paywalls — then falls back to a direct HTTP fetch (browser-like User-Agent, 10 s timeout, follows redirects, strips HTML to text), then to the Wayback Machine (web.archive.org).
3. Judge — one or more pluggable judges decide a verdict (Match / Partial / Mismatch / Uncertain / Inaccessible) + a one-line reason. Multiple judges = independent cross-check; an optional --tiebreak judge resolves disagreements.
4. Score — apply the 100-point rubric and assign a tier.
5. Report — emit <basename>_report.md (+ <basename>_claims.jsonl for reuse). Any tier-F citation raises a document-level warning banner.

Best fit

paper-verify works best as the first pass before human review:

• Research papers / reports — find the citations most likely to need manual paper reading (F, C, Uncertain, judge disagreement, weak author/year match, dead landing pages).
• Blog posts / newsletters — catch dead links, soft-404s, claim/source mismatch, and source drift before publishing.
• Lecture notes / public handouts — separate probably-safe citations from sources that need a human spot-check.
• Agent workflows — let Claude Code, Codex, Cursor, or Gemini parse JSON and loop only over the risky citations.

It is not a replacement for final expert review. It is designed to make that review smaller, faster, and better targeted.

Quickstart

Run it without installing anything (needs uv):

uvx paper-verify yourdoc.md --level L2 --out /tmp/pv

Runs with no API keys (keyword judge, low confidence). Output resembles:

paper-verify: 5 citations, level L2, judges: keyword
Overall: <score>/100 <tier>  [🟢A:<n> 🟡B:<n> 🟠C:<n> 🔴F:<n>]
⚠️  Document contains tier-F citations — see report.
Report:  /tmp/pv/yourdoc_report.md
Claims:  /tmp/pv/yourdoc_claims.jsonl

For a real fact-check, add an LLM judge:

export ANTHROPIC_API_KEY=sk-...
uvx --from "paper-verify[anthropic]" \
  paper-verify paper.md --level L2 --judge anthropic:claude-sonnet-4-6

💬 No terminal? Use it from a web chat

• Any web chat (Claude / ChatGPT / Gemini with browsing): copy-paste docs/webchat/webchat-prompt.md (한국어: webchat-prompt.ko.md) — the model fetches your sources and scores them with this same 100-point rubric.
• claude.ai (skill upload): upload the web-chat skill zip — extraction and scoring run as bundled code, fetching/judging use Claude's web tools, and the score comes from the real rubric via --from-evidence. Get the zip from Releases or build it: python tools/build_webchat_skill.py.

Install

pip install paper-verify                 # core, stdlib only — from PyPI
pip install "paper-verify[anthropic]"    # + Anthropic judge
# extras: [anthropic] [openai] [gemini] [mcp] [all] [dev]
pipx install paper-verify                # isolated CLI

# latest from source:
pip install "paper-verify @ git+https://github.com/nolainjin/paper-verify"

# for development:
git clone https://github.com/nolainjin/paper-verify && cd paper-verify
pip install -e ".[dev]"

Requires Python ≥ 3.10. From a clone you can also run without installing: python -m paperverify yourdoc.md --level L2.

Verification levels

L1 runs with no LLM at all. Pick the level with --level L1|L2|L3.

⚠️ L1 caveat. L1 scores reachability only (HTTP-alive → 100, unreachable → 0). It does not verify that the page content supports the claim. Soft-404s (pages that return HTTP 200 with error / placeholder content) are now detected heuristically — a reachable-but-suspect page scores 50 (not a clean 100), with a soft_404_suspect flag in the output. The heuristic is not perfect (it checks error markers, deep-path→homepage redirects, and suspiciously tiny bodies); use L2 / L3 for real content verification.

Academic metadata (paywall bypass)

For academic identifiers (DOI / arXiv / PMID / PMC, or URLs that clearly carry one), paper-verify queries free official metadata APIs before scraping HTML:

This returns structured title / authors / year / abstract even when the publisher landing page is a paywall stub, and makes the author/year rubric a real comparison instead of a fuzzy HTML match.

Explicit, observable fallback chain (each step only on failure of the prior; the path that actually served the data is recorded in the source field, never silently — per the No-Silent-Fallback principle):

metadata API (crossref|arxiv|ncbi)  →  HTTP fetch (http)  →  Wayback (archive)  →  none

• A metadata call uses a short timeout (~8 s) + one retry/backoff on transient errors (timeout / connection / HTTP 429 / 5xx); HTTP 404 means "not found" (no retry).
• A failed metadata lookup never crashes the run — it falls through, and source then reads "http" (so you can see the API did not serve it), "archive", or "none".
• If the whole chain fails, the citation is source="none", status=0, carries an error, and is scored Inaccessible — no invented metadata, never scored as alive.

The source field lets you see at a glance whether a citation was metadata-verified (crossref/arxiv/ncbi), HTML-scraped (http), served from archive, or unverifiable (none).

Judges & providers

Pass --judge SPEC (repeatable for cross-check). Spec forms:

The keyword judge is clearly low-confidence — it only measures lexical overlap, not meaning. Use it to smoke-test the pipeline; use an LLM judge for real verification.

Harness profiles (--profile)

A harness profile bundles the recommended judge order and frontend setup for a given agent (Claude Code, Cursor, Codex, Gemini). Pass --profile <key> and, when you do not pass any explicit --judge, paper-verify defaults the judges to that profile's recommended list, trying them in order of availability — it skips any judge whose SDK/CLI is not installed and uses the first available one(s), falling back to keyword (with a one-line stderr note) if none are available. An explicit --judge always wins; the profile is still recorded.

paper-verify paper.md --profile claude-code --json

Keys (aliases like claude → claude-code are accepted): claude-code, cursor, codex, gemini. The active profile is recorded in the JSON output under the top-level "profile" field (null when unset). List every profile as JSON (no file argument needed):

paper-verify --list-profiles | python -m json.tool

See docs/harness-strategy.md for the full matrix.

Agent packaging phases

Treat this repository as a staged agent integration:

Use Phase 2 when you want the workflow to be available inside an existing Codex setup without packaging a full plugin. Use Phase 3 when you want a distributable agent integration: plugin metadata, skill discovery, and MCP server wiring live together.

How cross-check works

Supply two or more --judge flags. Each judge evaluates the same (claim_context, source_text) independently. The verdict drives the claim-match score; the cross-check rubric item awards 10 points only when judges agree — so disagreement costs points and surfaces citations worth a human spot-check.

paper-verify paper.md \
  --judge anthropic:claude-sonnet-4-6 \
  --judge gemini:gemini-2.0-flash

Tie-break (3rd judge). When two or more judges disagree, pass an optional --tiebreak <spec> judge. It runs only on the split citations and restores the original 3-stage consensus spirit:

• on a genuine consensus (incl. after the tie-break) the citation keeps its 10 cross-check points, using the majority verdict as consensus (the tie-break judge arbitrates a genuine tie between distinct verdicts);
• if judges remain split with no --tiebreak, the effective verdict becomes Uncertain (claim-match = 15) and cross-check = 0, flagging it for review.

paper-verify paper.md \
  --judge anthropic:claude-sonnet-4-6 \
  --judge openai:gpt-4o-mini \
  --tiebreak gemini:gemini-2.0-flash

A judge may also answer Uncertain on its own when the source is insufficient to decide — better than guessing. Uncertain citations are grouped in a "Needs re-check" section of the Markdown report.

Using paper-verify from your own agent

paper-verify is agent-callable two ways. Both are provider-agnostic — pick any judge (keyword, anthropic, openai, gemini, cli:*); the structured output shape is identical regardless of judge.

Provider-specific harness profiles are documented in docs/harness-strategy.md, with frontend notes for Claude Code, Cursor, Codex, and Gemini under docs/providers/. The core pipeline stays shared; frontend differences live in profiles and docs instead of long-lived provider branches.

(a) --json — capture structured output from stdout

Add --json and the CLI writes the full result as JSON to stdout while the human summary goes to stderr, so an agent can capture and parse stdout directly:

result=$(paper-verify paper.md --level L2 --judge keyword --json)
echo "$result" | python -m json.tool
# overall_score / overall_tier / has_failure live at the top level:
echo "$result" | python -c "import sys,json; d=json.load(sys.stdin); print(d['overall_score'], d['overall_tier'], d['has_failure'])"

The JSON top-level keys are: schema_version, source_file, level, profile (active harness profile key, or null), judges, overall_score, overall_tier, has_failure, tier_distribution (counts per tier), and citations (one object per citation: citation, fetched, judgements, consensus, effective_verdict, score, breakdown, tier).

Each citation.fetched object carries (schema_version "4"): status, title, abstract, url_final, via_archive, error, plus authors (list, from metadata APIs), year (int or null), source (crossref | arxiv | ncbi | http | archive | none — which path produced the data), and soft_404_suspect (bool — a 2xx page that looks like an error/placeholder).

For triage automation, prioritize citations where any of these are true:

• tier is F or C.
• consensus / effective_verdict is Uncertain, Mismatch, or Inaccessible.
• judges disagree in judgements.
• fetched.soft_404_suspect is true.
• fetched.landing_status is 403, 404, or another non-2xx status while metadata still resolved the source.
• fetched.source is archive or none.

--json is additive: pass --out DIR to also write the .md / .jsonl files. Exit codes are stable for agents: 0 = ran successfully regardless of grades (a tier-F document still exits 0 — inspect has_failure); a nonzero code (2) means a real error (file not found, bad judge spec).

(b) MCP server — register as a tool

paper-verify[mcp] ships an MCP server (stdio transport) exposing these tools:

pip install "paper-verify[mcp]"

Register it with an MCP client. For Claude Code:

claude mcp add paper-verify -- paper-verify-mcp

Or in an MCP client config (mcpServers):

{
  "mcpServers": {
    "paper-verify": {
      "command": "paper-verify-mcp"
    }
  }
}

The mcp package is an optional extra — the core tool and --json work with mcp not installed; importing the server without it raises a clear install hint for the [mcp] extra instead of crashing.

(c) --from-evidence — bring your own fetch/judge

If your agent (or a web chat) already fetched the sources and judged the claims, hand paper-verify the evidence and let it apply the standard rubric — identical scoring to a native run:

paper-verify yourdoc.md --extract-only > citations.json   # deterministic extraction
# …your agent fetches each citation and judges it, producing evidence.json…
paper-verify --from-evidence evidence.json --json --out out/

The evidence shape is documented by example in examples/evidence-sample.json: per citation, the citation object from --extract-only verbatim, a fetched object (status, title, abstract, authors, year, source, soft_404_suspect, …), and one or more judgements ({"judge", "verdict", "reason"} — verdicts: Match | Partial | Mismatch | Uncertain | Inaccessible). Malformed evidence exits 2 with the offending citation index named. This is the engine behind the web-chat skill.

Scoring rubric (100 points)

At L1, a reachable page scores 100, a soft-404-suspect reachable page scores 50, and an unreachable page scores 0.

Document score = average of per-citation scores. If any citation is tier F, the whole document is flagged ⚠️.

Verdict & tier taxonomy

Verdicts (per judge):

Tiers (per citation, and document average):

Limitations

• AI judgment reliability — LLM judges can mis-read an abstract's meaning. Cross-check with a second judge (+ --tiebreak) and spot-check important citations by hand. A judge may answer Uncertain rather than guess.
• Paywalls — for DOI / arXiv / PMID / PMC the free metadata APIs return the title / authors / year / abstract even behind a paywall. Full-text claims that need the body (not just the abstract) may still land in tier C — expected.
• Soft-404 detection is heuristic — it catches common error markers, deep path→homepage redirects, and tiny bodies, but can miss disguised error pages.
• Archive misses — when a URL is dead and absent from the Wayback Machine, the fallback fails and the citation is marked Inaccessible.
• JavaScript-heavy pages — SPA pages may return little text to the stripper.
• The keyword judge measures lexical overlap only and is not a substitute for semantic verification.

한국어 요약

문서 안의 인용 출처(URL·DOI·PMC·PMID·arXiv)를 추출해 실제 원문과 대조하고 100점 루브릭으로 채점하는 도구입니다. 전체 한국어 가이드: README.ko.md (터미널 없이 웹챗에서 쓰는 방법 포함).

License

MIT © 2026 Duchan Jin (진두찬). See LICENSE.