---
name: cited-source-mirror-verification
description: "Verify a cited empirical/numeric claim against an open mirror (aiXiv, OpenAlex, Semantic Scholar, the author's own copy) before citing it. Use whenever you are about to cite a statistic, effect size, odds ratio, percentage, p-value, or empirical result you have NOT read in the primary — because the primary is access-blocked (SSRN / publisher 403, paywalled PDF), or because the claim arrived via an LLM research digest, daily-research note, or WebSearch summary (which routinely reframe a methodological artifact or a limitation as a headline finding). Especially before committing citations to a durable artifact — paper, knowledge graph, .zenodo.json, ADR, DOI repo."
user-invocable: false
origin: shimo4228
---

# Verify Blocked Primary Sources Against an Open Mirror

A number you cite carries the authority of its primary, but it usually reaches
you through a chain — **LLM-generated report → WebSearch summary → secondary
blog → primary paper** — and each hop can drop or distort the load-bearing
detail. The primary is also frequently access-blocked (SSRN and many publishers
return HTTP 403 to bots; paywalled PDFs). Citing from the digest ships the
secondary's framing, not the source's. This skill is the discipline that closes
that gap before a citation becomes durable.

## Why this matters (the failure it prevents)

Concrete case (2026-06): a research digest cited *"a significant negative
association (OR = 0.546)"* and a WebSearch summary presented it as a finding —
"schema markup reduces AI citation." The primary abstract, reached only through
an open mirror (aiXiv), flagged that same number explicitly as a **methodological
artifact** that the controlled analysis then corrected for. The real finding was
the opposite-shaped *"presence isn't the lever; richness is."* Citing the
secondary framing would have shipped a misleading claim into a DOI-registered
repository whose own stated discipline is citation fidelity — a self-inflicted
ghost citation.

The lesson is not "digests are bad." It is that **a summary preserves the number
while losing its epistemic status** (finding vs artifact vs baseline vs
limitation), and the status is exactly what makes the citation true or false.

## How to verify

When the claim is empirical/numeric AND the primary is access-blocked (or the
claim came via a digest you have not read past):

1. **Find an open mirror of the *same* paper**, in rough order of fidelity:
   - aiXiv / arXiv — SSRN and conference papers are often cross-posted; search by title.
   - OpenAlex (`https://api.openalex.org/works/doi:<DOI>`) or Semantic Scholar —
     but confirm the abstract field isn't null before trusting it.
   - The author's own copy / institutional page / a slide deck.
   - Google Scholar's indexed snippet — last resort, partial, but often quotes the abstract.
2. **Verify against the mirror, not the digest** — three things, not just the number:
   the **exact value**, its **direction** (which arm/group is higher — summaries
   swap these), and its **framing** (is the headline figure a finding, an artifact,
   a baseline, or a limitation the paper itself rejects?).
3. **Treat the digest/summary as a lead, never a source.** Trust degrades up the
   chain: `primary > open mirror of the primary > OpenAlex/Scholar > secondary blog > LLM digest`.
   A WebSearch result that quotes a marketing blog is two hops below the paper.
4. **If nothing openable verifies the numbers, cite only what you can.** Soften to
   the verifiable core (the controlled design, the qualitative direction) and drop
   the specific figures rather than passing through an unverified number. State
   that the specifics are unverified if they matter.

## Worked check

```bash
# 1. Is the primary actually blocked?
curl -s -o /dev/null -w "%{http_code}\n" "https://www.ssrn.com/abstract=<id>"   # 403 = blocked

# 2. Resolve the record — title search (works with only a half-remembered title):
curl -s "https://api.openalex.org/works?search=<title+words>&per-page=3&mailto=<your-email>" \
  | python3 -c "import sys,json;d=json.load(sys.stdin);[print(w['display_name'],'|',w.get('doi'),'| OA:',w['open_access']['oa_url']) for w in d['results']]"

# ...or straight from a DOI you already have:
curl -s "https://api.openalex.org/works/doi:<DOI>?mailto=<your-email>" \
  | python3 -c "import sys,json;d=json.load(sys.stdin);print(d['display_name']);print(d['open_access']['oa_url'])"
```

**`oa_url` が blocked な primary に戻ることは普通にある** — OpenAlex は *レコード* を
確定させる層であって、必ずしも読める全文をくれるわけではない。実測 (2026-07-25、上の
schema-markup ケース) では title search が 1 位でこの論文を返したが `oa_url` は SSRN
DOI そのものだった。**そこから先は読める mirror を探す別ステップ**になる:

- WebSearch で `"<exact title>" aixiv OR arxiv` — 上の事例で実際に効いたのはこれ (aiXiv)
- 著者の個人ページ / 所属機関リポジトリ / スライド
- Semantic Scholar API も候補だが、**API key 無しでは 429 を返す**ことが多く当てにしない

見つけたら WebFetch で abstract ページを取り、こう問う: *exact value は何か、どちらの
群が高いか、その数値を論文は finding として出しているか artifact / limitation として
出しているか* — 数字だけでなく **epistemic status** を確定させる。

## When this triggers

- You are about to cite a statistic, effect size, OR/coefficient, percentage, or
  result that you have **not read in the primary**.
- The claim arrived through an **LLM digest, daily-research note, or WebSearch summary**.
- The primary returns **403 / is paywalled / has no mirrored abstract**.
- You are committing citations to a **durable artifact** — a paper, a
  `graph.jsonld` / `.zenodo.json` citation surface, an ADR, a DOI release.
- A source-fidelity reviewer reports a cited number as **UNCHECKED** because the
  primary could not be opened.

## Anti-patterns

- Citing a number straight from an LLM digest or a single SEO/marketing blog.
- Trusting a WebSearch summary's *framing* — it may have silently dropped
  "this was an artifact" or "this did not replicate."
- **Chain-of-secondaries**: your repo's files all agree with one another because
  they share one unverified upstream. That is internal consistency, not
  verification — every surface being wrong the same way looks identical to every
  surface being right.

## Related

- `fact-checker` agent → "Local-Source Verification" — the complement. That mode
  reconciles a *draft's internal* claims (dates, counts, order) against machine
  records (logs, git, timestamps); this skill verifies *external cited primaries*.
  Reach for that one for "did this happen when/how-many-times I wrote," and this
  one for "is this cited number real and framed correctly."