citation-context

Citation Context

Analyze whether citations are supportive, contrastive, or merely mentioning. Helps build stronger arguments by ensuring citation framing matches the actual relationship between papers.

Citation Role Taxonomy

Supporting

The cited work provides evidence for or agrees with the current claim.

• Signal phrases: "As shown by", "Consistent with", "In line with", "Confirmed by", "Supporting this", "corroborates"
• Example: "As demonstrated by Smith et al. (2023), attention mechanisms improve performance on long-range dependencies."

Contrasting

The cited work disagrees with, contradicts, or presents an alternative to the current claim.

• Signal phrases: "Unlike", "However", "In contrast to", "Contradicting", "While X found", "disputes"
• Example: "Unlike Jones (2022), who reported no significant effect, our results show a clear improvement."

Mentioning

Neutral reference without evaluative framing. Acknowledges existence or provides background.

• Signal phrases: "X studied", "X proposed", "X introduced", "X reported", "According to"
• Example: "Smith (2023) proposed a transformer-based approach for protein folding."

Extending

The current work builds directly on the cited work, adding to it or adapting it.

• Signal phrases: "Building on", "Extending the work of", "Adapting the approach of", "Inspired by"
• Example: "Building on the framework of Smith (2023), we introduce a multi-scale attention mechanism."

Methodological

The cited work provides a method, tool, dataset, or metric used in the current study.

• Signal phrases: "Following the approach of", "Using the method described in", "Evaluated on the benchmark of", "Implemented using"
• Example: "We evaluate our model on the benchmark introduced by Smith et al. (2023)."

Deterministic backend

Citation enumeration and every existence claim about a citing or cited paper route through the researcher-core evidence kernel. Claude keeps the interpretive work (deciding whether a citation is supporting, contrasting, mentioning, extending, or methodological); core does the reproducible work (enumerating who cites a paper, resolving four-state identity, and checking retraction status). Full command reference: references/core-cli.md.

Enumerate who cites a paper (citation graph)

uv run --project "${CLAUDE_PLUGIN_ROOT}/core" python -m researcher_core citations <doi-or-id> --limit 50 --json
uv run --project "${CLAUDE_PLUGIN_ROOT}/core" python -m researcher_core references <doi-or-id> --limit 50 --json

• citations <id> returns forward citations (works that cite this one); references <id> returns backward references (works this one cites). Both return {seeds[], direction, depth, nodes[], edges[], warnings[], sources[], counts{}}, and each nodes[] element is a CSL-JSON record. Read nodes[] for the enumerated works and warnings[] for any source that failed.
• Default sources are openalex,semantic_scholar,opencitations. A source in warnings[] is a coverage gap, not evidence a paper is uncited.

Verify any citing entry asserted as existing (axis a, four-state)

uv run --project "${CLAUDE_PLUGIN_ROOT}/core" python -m researcher_core verify-ref "<doi-or-title>" --json

Read, per entry: verdict (verified / mismatch / unresolvable / inconclusive), refusal_grade (already computed, never re-derive it), reason, source_outcomes[], best_match, and the status and accessibility blocks. Consume the four-state verdict, NEVER a boolean.

• Refusal-grade fires ONLY on unresolvable or mismatch (plus retracted from axis b, contradicted from axis c). A citing entry that is unresolvable or mismatch is FLAGGED as unresolved, never listed as an established citing work.
• inconclusive is NEVER refusal-grade. It means a source errored or only one index holds the paper. Surface it as an open item ("could not confirm, one index only"), never as fabrication. Accusing an honest author of inventing a real citation is the worst failure here.

Retraction annotations on citing papers (axis b)

uv run --project "${CLAUDE_PLUGIN_ROOT}/core" python -m researcher_core status <doi-or-id> --json

Read entries[].status for current / corrected / retracted / expression-of-concern. Annotate any citing paper that is retracted or expression-of-concern inline: a retracted work still counts in the tally but must carry the notice. status.checked: false means axis (b) was never answered, which is absence of evidence, not a clean bill of health.

Framing faithfulness (axis c), and the layer it names

When auditing whether the user's framing matches the cited paper's actual finding, anchor on M2 passages:

uv run --project "${CLAUDE_PLUGIN_ROOT}/core" python -m researcher_core passages index <doi-or-url> --json
uv run --project "${CLAUDE_PLUGIN_ROOT}/core" python -m researcher_core faithfulness "<the user's claim>" --doc <doc-id> --json

• Every faithfulness verdict NAMES its layer: full-text when an OA copy was indexed, abstract-only otherwise. When full text is unavailable the verdict is insufficient-passage, never "clean" or "faithful"; surface it as an open item, never refusal-grade.
• Axis (c) is a LEXICAL baseline (BM25 plus token-overlap and polarity heuristics). It can miss overstatements: a framing it calls supported may still overreach. Treat supported as "no lexical contradiction found at this layer", not proof the framing is accurate. Only contradicted is refusal-grade on this axis.

Scite and Claude, side by side

When the Scite MCP connector is connected, its tallies add quantitative supporting / contrasting / mentioning counts across the whole literature. These COMPLEMENT, never replace, Claude's interpretive classification of the specific sentence in front of you. Report both: the Scite tally (where present) and Claude's role judgment for this manuscript's usage.

Degradation path (state it in the output when a fallback is used)

1. Core present (uv + core/, or pip install -e core/): citation graph, four-state identity, status, and faithfulness as above.
2. No uv / no core: fall back to the Scite MCP (smart citations, tallies, editorial notices) and Zotero MCP (library metadata) where connected.
3. No MCP either: web search plus local \cite text analysis only. Say so, and never assert a citing work exists or a paper is retracted without a retrieval behind it.

Operations

Enumerate Citing Works

Triggered by: "who cites this paper", "analyze citations of this paper", "citation audit" for a specific paper

Workflow:

1. Resolve the target paper (DOI, arXiv ID, or OpenAlex ID).
2. Run core citations <id> --json to enumerate the works that cite it (and references <id> for what it cites). Read nodes[]; note any source in warnings[] as a coverage gap.
3. For each citing work you assert as existing, run verify-ref and consume the four-state verdict. List verified and inconclusive works (the latter flagged "one index only"); FLAG unresolvable / mismatch works as unresolved, never as established citations.
4. Run status over the citing set; annotate any retracted or expression-of-concern paper.
5. If Scite is connected, attach its supporting / contrasting / mentioning tallies. Claude classifies the interpretive role of each citation's framing.

Output:

Citing Works for Smith et al. (2023)  [DOI 10.xxxx/yyyy]
========================================================
Enumerated: 42 citing works (openalex 42, semantic_scholar 38, opencitations 40)
Coverage note: none (all sources returned)

Verified citing works:            39
Inconclusive (one index, open):    2
Unresolvable (flagged, NOT fact):  1

Retraction notices:
  jones2021: RETRACTED (axis b), still tallied, notice attached

Scite tally (where available): supporting 12, contrasting 5, mentioning 25
Claude role read (this manuscript's usage): 3 supporting, 1 contrasting

Analyze Citation Roles in Manuscript

Triggered by: "analyze my citations", "citation roles", "how am I using my citations"

Workflow:

1. Parse all .tex files in manuscript/ for citation commands (\cite, \citep, \citet, \parencite, \textcite)
2. For each citation, extract the surrounding sentence and paragraph context
3. Classify the citation role based on signal phrases and semantic context
4. Generate a citation context report

Output:

Citation Context Report
=======================
Total citations analyzed: 47

Role Distribution:
  Supporting:     18 (38%)
  Contrasting:     7 (15%)
  Mentioning:     12 (26%)
  Extending:       4  (9%)
  Methodological:  6 (13%)

Detailed Analysis:
  introduction.tex:14, \cite{smith2023}: Supporting
    Context: "As shown by Smith et al. (2023), transformers outperform..."
    
  methods.tex:31, \cite{jones2022benchmark}: Methodological
    Context: "We evaluate on the benchmark introduced by Jones (2022)..."

Audit Citation Framing

Triggered by: "audit citations", "check citation framing", "are my citations accurate"

Workflow:

1. For each citation in the manuscript, identify the role (as above).
2. Confirm the cited paper exists with core verify-ref; consume the four-state verdict and attach any status retraction notice. Flag unresolvable / mismatch entries rather than auditing framing against a paper that may not exist.
3. Anchor the user's framing against the source with core faithfulness (see Deterministic backend). Every verdict names its layer (abstract vs full-text); an insufficient-passage verdict is an open item, not a mismatch. Axis (c) is a lexical baseline that can miss overstatements, so a supported verdict is not proof the framing is accurate.
4. If Scite MCP is connected, add its smart-citation tallies as corroboration.
5. Flag genuine framing mismatches (a contradicted faithfulness verdict, or a Scite tally that inverts the user's stated role); the interpretive call stays with Claude.

Output:

Citation Framing Audit
======================
Potential issues found: 3

WARNING: introduction.tex:22, \cite{smith2023} framed as Supporting
  Your text: "Smith et al. (2023) confirmed that X improves Y"
  Actual finding: Smith et al. found X improves Y only under condition Z
  Suggestion: Add qualifier: "under condition Z"

WARNING: discussion.tex:45, \cite{jones2022} framed as Contrasting
  Your text: "Unlike Jones (2022), we found..."
  Actual finding: Jones reported mixed results, not a clear contradiction
  Suggestion: Reframe as "Extending the mixed findings of Jones (2022)..."

Suggest Citation Framing

Triggered by: "how should I cite this", "frame this citation", "cite this as"

Workflow:

1. User provides a citation key and the claim they want to support.
2. Retrieve the cited paper via core get (abstract and metadata) or fulltext (OA full text where available); fall back to Scite MCP or literature-search if core is absent.
3. Determine the actual relationship with core faithfulness, naming the layer verified (abstract vs full-text). Remember axis (c) is a lexical baseline and can miss overstatements.
4. Suggest appropriate framing with example sentences.
5. Warn if the paper does not actually support the intended use. When the verdict is insufficient-passage (no OA full text), say the framing could not be checked at full-text level, and surface it as an open item rather than endorsing the framing.

Output:

Citation Framing Suggestion for \cite{smith2023}
================================================
Your claim: "Attention mechanisms are essential for long-range dependencies"
Paper finding: Smith et al. showed attention helps but is not strictly necessary

Recommended role: Supporting (with qualification)
Suggested framing:
  "Smith et al. (2023) demonstrated that attention mechanisms significantly
   improve performance on long-range dependency tasks, though alternative
   approaches such as state-space models show competitive results."

NOT recommended:
  "Smith et al. (2023) proved that attention is essential..." (overstates finding)

Build Argument Support Map

Triggered by: "map citations to arguments", "citation support map", "which citations support which claims"

Workflow:

1. Extract the argument structure from the manuscript (or from an argument-map.md if available)
2. Map each citation to the claim(s) it supports or challenges
3. Identify claims with insufficient citation support
4. Identify citations that are not clearly linked to any claim
5. Suggest additional citations needed for under-supported claims

Output:

Argument Support Map
====================

Claim: "Method X outperforms baselines"
  Supporting: smith2023, jones2024, chen2023  (3 citations)
  Contrasting: wang2022  (1 citation, addressed in discussion)
  Status: Well-supported

Claim: "Training is more efficient than prior approaches"
  Supporting: smith2023  (1 citation)
  Contrasting: none found
  Status: Under-supported: consider additional evidence

Scite MCP Integration

Scite complements the deterministic backend; it does not replace it. Core enumerates citing works and resolves identity and status deterministically (see Deterministic backend); Scite adds literature-wide quantitative tallies when connected:

• "Smart citations": actual excerpts showing how papers cite each other
• Citation statement classifications (supporting, contrasting, mentioning) as counts
• Cross-references across the literature, not just the user's manuscript
• Editorial notices for retractions or corrections (corroborating core's axis-b status)

The interpretive role classification for a specific sentence always stays with Claude. When neither core nor Scite is available, fall back to local text analysis of the user's manuscript (signal-phrase role detection, abstract-level comparison via literature-search) and say so.

Integration Points

• literature-search: Retrieves paper metadata and abstracts for framing analysis
• citation-management: Shares library.bib entries; may flag entries needing framing review
• research-convergence: Argument maps inform which citations should be supporting vs contrasting
• paper-drafting: Citation context analysis guides how to frame references during drafting
• peer-review: Reviewers check whether citation framing is accurate and fair
• post-draft-integrity hook: Can trigger citation framing audit after drafting completes

Integrity constraints

• Never fabricate citations: every reference must come from an actual retrieval (API, MCP, or user-provided source). If a citation cannot be verified, flag it; never invent a DOI, author list, venue, or year.
• Never invent data: only user-provided or actually computed numbers may appear as results. Anything illustrative must be labeled "(synthetic, for demonstration)".
• Refuse to present as valid output: a likely-fabricated or unresolvable citation, a data claim with no traceable source, or a retracted source (unless the user explicitly cites it as retracted).
• Consume the four-state identity verdict, never a boolean. Refusal-grade behavior (flag as unresolved, withhold from the citing list) fires ONLY on unresolvable or mismatch on axis (a), retracted on axis (b), or contradicted on axis (c). inconclusive and insufficient-passage are NEVER refusal-grade: surface them as open items. Treating inconclusive as fabrication accuses an honest author of inventing a real citation.

Canonical copy: references/integrity-constraints.md.