huguryildiz lit-anchor
Every citation is a claim that a specific work exists and says a specific thing. AI assistants produce plausible-but-fabricated references — real-looking DOIs that resolve to nothing, or to a different paper; correct-sounding titles with invented authors or years. A bulk "give me twenty references" response is a draft to verify, never a result. This skill is the discipline that keeps a fabricated citation out of your paper.
When to use
- Assembling or extending a bibliography / related-work section.
- Adding a single citation to support a claim.
- Making a novelty or absence claim ("no prior work does X", "we are the first to…").
- Any moment an AI hands you a formatted, confident-looking citation.
When NOT to use
- Summarizing or reasoning about a paper you already hold — that is reading, not sourcing.
- Internal cross-references to your own artifacts (use
layer-sync/decision-log).
Runbook
- Treat every AI-produced identifier as a hypothesis. A DOI, title, author list, year, or venue emitted by a model is unverified until resolved against an authoritative source.
- Verify one citation at a time. Resolve each DOI/URL against a real source (the DOI resolver, the publisher page, an indexed search) and confirm title + authors + year + venue all match. Do not batch — batching is where fabrications hide.
- Main agent synthesizes; sub-agents fetch small pieces. Never delegate "produce the verified bibliography" to a sub-agent that returns a finished, verified-looking table — that is exactly the pattern that launders fabrications. Sub-agents retrieve single documents; the main agent assembles and vouches.
- Mark unverified and abstract-only sources explicitly. An abstract-only source carries a documented "full text before submission" debt; an unresolved citation is flagged, never silently included.
- Back absence claims with a documented search. "No prior work does X" requires the actual queries run and their date recorded — not a model's confidence.
Rules
- No fabricated identifiers. Every DOI/URL in the record resolves to the exact claimed work.
- Bulk AI citation output is a draft. It enters the record only after per-item verification.
- Every citation carries provenance: where it was verified and the as-of date; trained knowledge about a paper is a hypothesis about the present, not an observation.
- Absence/novelty claims are search-backed, with the search recorded.
- A citation that will not resolve after a documented search is dropped or flagged — never quietly kept because it "looks right".
Configuration
{{bibliography}}— where verified citations live (e.g.docs/references.bib, a notes file).{{claim_log}}— optional: where novelty/absence claims and their backing searches are recorded.
Provenance & maintenance
Generalized from a working computational-research repository's DOI-verification protocol, adopted
after a bulk literature-review sub-agent was caught emitting fabricated references: the fix was
small-step manual retrieval with one-by-one resolution and main-agent synthesis, plus an
explicit abstract-only debt on sources not yet read in full. See the pack's examples/ directory
for the worked exemplar mapping.
Re-verify in your project:
- Resolve every identifier in
{{bibliography}}and confirm the resolved metadata matches the entry (a stale or redirected DOI is a silent corruption). - Confirm each novelty/absence claim in
{{claim_log}}still has its backing search recorded.