Literature Review

When to Use

• Starting a new project and need a scoped set of relevant papers
• Checking whether a finding has prior literature
• Producing a bibliography for a draft or proposal
• Finding citations to back specific claims already written in a draft

Decomposing Claims Before Searching

When the input is a set of claims to support — a draft paragraph, an introduction, or a brainstorm decomposition tree (brainstorm.md) — do not search the claim wholesale. A single sentence usually bundles a fact and an interpretation, and each needs a different kind of source. Split every claim into:

• Factual statements — empirical or descriptive assertions (“adoption rose 30% since 2019”, “the median firm has 12 employees”). Each needs a citation that establishes the fact: a dataset paper, an empirical study, or an official statistic.
• Argumentative statements — interpretive, causal, or theoretical positions (“X raises Y because Z”, “this effect reflects status competition”). Each needs literature that supports the mechanism or position.

Run one search per statement (each feeds Path A or Path B below), then attach the verified citations back to the parent claim so the draft can cite each statement precisely. The per-statement citations feed paper-writing.

Recency over foundational. For every statement, prefer recent literature (within 5 years) over foundational high-impact papers. Cite at most one foundational/seminal paper per argument — the recent work carries the support; the single foundational paper anchors lineage only. Stacking multiple seminal papers on one argument is padding, not support.

Path A — API-Grounded (higher precision)

OpenAlex search → relevance triage → Crossref for full metadata + DOI

Step 1 — Search OpenAlex

import requests

def search_openalex(query, n=50):
    url = "https://api.openalex.org/works"
    params = {
        "search": query,
        "per-page": n,
        "sort": "relevance_score:desc",
        "mailto": "you@example.com",   # polite pool — faster responses
    }
    return requests.get(url, params=params).json()["results"]

Key fields returned: title, doi, publication_year, authorships, cited_by_count, open_access.oa_url.

Step 2 — Triage by relevance

Score each result against the research question. Keep papers where title + abstract clearly match. Discard tangential hits.

Step 3 — Verify via Crossref

def crossref_lookup(doi):
    r = requests.get(f"https://api.crossref.org/works/{doi}",
                     headers={"User-Agent": "ResearchBot/1.0 (mailto:you@example.com)"})
    if r.status_code != 200:
        return None
    w = r.json()["message"]
    return {
        "title":   w.get("title", [""])[0],
        "authors": [a["family"] for a in w.get("author", [])],
        "year":    w.get("published", {}).get("date-parts", [[None]])[0][0],
        "journal": w.get("container-title", [""])[0],
        "doi":     w.get("DOI"),
    }

Cross-check title, authors, year, and venue against the OpenAlex record. Flag mismatches.

Path B — Verify an Existing Bibliography

Parse input → Crossref DOI lookup → title-search fallback → classify → report

Use when a list of citations already exists and needs to be audited before it can be trusted:

• Web search produced a draft bibliography → verify before using
• A paper’s reference section needs spot-checking
• Another LLM was asked for sources and the output is suspicious
• An old bibliography needs re-checking for broken DOIs
• Topic is niche enough that OpenAlex coverage is sparse and grey literature (working papers, reports) is in the mix

Non-negotiable rules

1. Every entry is classified into exactly one of: VERIFIED, MISMATCH, FABRICATED.
2. Never “fix” a FABRICATED entry by substituting a similar-looking real paper. Flag it and stop. The job is to report, not to repair — the user decides what to do.
3. Never downgrade a MISMATCH to VERIFIED because “it is close enough”. Record exactly what was wrong.

Step 1 — Parse the input

Accept any common format: APA, BibTeX, Chicago, plain text, or a markdown list. For each entry, extract:

• First author surname
• Year
• Title
• Venue (journal / conference name)
• DOI (if present)

If a field is missing, record it as null and proceed — missing fields are informative.

Step 2 — Resolve DOIs at Crossref

For entries that have a DOI:

curl -s "https://api.crossref.org/works/$DOI" > crossref.json

Parse the JSON and check:

• First author surname: exact match (case-insensitive) against Crossref’s author[0].family.
• Year: exact match against issued.date-parts[0][0].
• Title: ≥ 80% fuzzy match against title[0] (use difflib.SequenceMatcher from Python stdlib).

Record which fields passed and which failed.

Step 3 — Title-search fallback for entries with no DOI

For entries with no DOI (or where the DOI returned HTTP 404), search Crossref by title:

ENCODED_TITLE=$(python3 -c "import urllib.parse,sys; print(urllib.parse.quote(sys.argv[1]))" "$TITLE")
curl -s "https://api.crossref.org/works?query.title=$ENCODED_TITLE&rows=3" > fallback.json

For each of the top 3 hits, compare first author and year against the claimed citation. If any hit matches both, record its DOI as evidence.

Step 4 — Classify every entry

Apply these rules in order:

A MISMATCH is not a fabrication. The paper is real — the citation just has errors (typos, misremembered years). The user may want to fix them.

Step 5 — Report

Produce a markdown file with a summary header and a verdict table:

# Citation Verification Report

**Input:** <source file or description>  |  **Entries audited:** N
**Verified:** V  |  **Mismatch:** M  |  **Fabricated:** F

| # | Original citation | Verdict | Evidence |
|---|---|---|---|
| 1 | Author, A. (2023). Title. *Venue*. | VERIFIED | https://doi.org/... |
| 2 | Author, B. (2022). Title. *Venue*. | MISMATCH | year should be 2021; https://doi.org/... |
| 3 | Author, C. (2024). Fake Title. *Venue*. | FABRICATED | no Crossref match for title or author+year |

## Recommended actions

- **Verified (V):** safe to use as-is.
- **Mismatch (M):** real papers, but fix the flagged fields before citing.
- **Fabricated (F):** drop from the bibliography. Do not cite.

For FABRICATED entries, the output is “FABRICATED — not found in Crossref”. Do not suggest “a similar paper” or “did you mean…”. The user decides.

Hallucination Taxonomy

Scope Defaults

Unless the user specifies otherwise, apply these filters:

Publication year: last 5 years (current year − 4 through current year). Extend to 10 years only if fewer than 10 relevant papers are found within 5.

Recency over foundational. Prefer recent literature (within 5 years) over foundational high-impact papers. When backing a specific argument, cite at most one foundational/seminal paper — the recent work carries the support; the single foundational paper anchors lineage only. Stacking multiple seminal papers on one argument is padding, not support. (See “Decomposing Claims Before Searching” above.)

Journal and venue whitelist:

Apply the whitelist during triage (Step 2), not during search. Cast a wide net in Step 1, then filter to whitelisted venues before verification.

OpenAlex venue filter (API-grounded path):

def search_openalex_scoped(query, n=100, year_from=None):
    import datetime
    if year_from is None:
        year_from = datetime.date.today().year - 4
    params = {
        "search": query,
        "per-page": n,
        "sort": "relevance_score:desc",
        "filter": f"publication_year:>{year_from - 1}",
        "mailto": "you@example.com",
    }
    results = requests.get("https://api.openalex.org/works", params=params).json()["results"]
    # Triage: keep whitelisted venues only
    whitelist = {
        # Econ top 5
        "american economic review", "quarterly journal of economics",
        "journal of political economy", "review of economic studies", "econometrica",
        # Econ field
        "american economic journal: applied economics", "american economic journal: economic policy",
        "american economic journal: macroeconomics", "american economic journal: microeconomics",
        "the economic journal", "rand journal of economics", "review of economics and statistics",
        # Marketing / IS
        "marketing science", "journal of marketing research", "journal of marketing",
        "management science", "mis quarterly", "information systems research",
        "quantitative marketing and economics",
        # Multidisciplinary science
        "proceedings of the national academy of sciences", "science", "nature",
        "nature communications", "science advances", "nature human behaviour",
        # CS / ML
        "neurips", "icml", "iclr", "acl", "emnlp",
        # CS / IR & Web
        "sigir", "the web conference", "recsys", "knowledge discovery and data mining",
    }
    return [r for r in results
            if any(w in (r.get("primary_location") or {}).get("source", {}).get("display_name", "").lower()
                   for w in whitelist)]

Stopping Criterion

If zero papers survive triage or verification, stop and report that honestly.

Do not fabricate entries to fill the bibliography. Return a short note such as:

No papers matching the research question were found in whitelisted venues within the target date range. Consider broadening the date window, relaxing the venue filter, or switching to Path B for grey literature.

This applies at every filtering stage: after venue triage, after relevance triage, and after verification.

Trade-offs

Report

Output uses the Quick Template — three labeled lines, Definition / Description / Takeaway. (For multi-section writeups, see report.)

Definition (measure): N papers found; N retained after triage; N verified clean; hallucination rate (unverified / total).
Analyses: Path used (API / web-search); databases searched (OpenAlex, Crossref, Google Scholar); verification method applied.
Takeaway: Coverage verdict for the research question; any fabricated or low-confidence entries flagged for human review.