Citation Audit

Citation Audit

🔒 Do not wrap this skill in /loop, /schedule, or CronCreate. It is verdict-bearing — it judges bibliographic correctness. Re-running that verdict on a timer adds no new signal (it changes only when the bibliography changes). Schedule the external wait that precedes it — bibliography finalized → then audit once. See shared-references/external-cadence.md.

Verify every \cite{...} in a paper against three independent layers:

1. Existence — the cited paper actually exists at the claimed arXiv ID / DOI / venue.
2. Metadata correctness — author names, year, venue, and title match canonical sources (DBLP, arXiv, ACL Anthology, Nature, OpenReview, etc.).
3. Context appropriateness — the cited paper actually supports the claim it is being used to support in the manuscript.

This skill is the fourth layer of \aris{}'s evidence-and-claim assurance, complementing experiment-audit (code), result-to-claim (science verdict), and paper-claim-audit (numerical claims). Together they form a bottom-up integrity stack from raw evaluation code to manuscript bibliography.

When to Use This Skill

Run before submission. The right gating point is:

• After paper-write has produced the LaTeX draft and bib file
• After paper-claim-audit has verified numerical claims
• Before final paper-compile for submission

Do not run this on a half-written draft — most of the work is in cross-checking each \cite against context, which is wasted on placeholder text.

What This Skill Catches

The dangerous citation problems are not wildly fake citations — those are easy to spot. The dangerous ones are:

• Wrong-context citations: real paper, but the cited claim is not what that paper actually establishes (e.g., citing Self-Refine to support "self-feedback produces correlated errors" — Self-Refine actually argues the opposite).
• Author hallucinations: anonymous-author placeholders that slipped through, missing co-authors, wrong order.
• Title drift: arXiv v1 vs v3 with different titles silently merged.
• Venue confusion: arXiv preprint cited but the official venue is now CVPR/ICML/NeurIPS — using the wrong record.
• Year mismatch: arXiv 2023 preprint with 2024 conference acceptance, year reported inconsistently.
• Phantom DOIs: DOI looks real but does not resolve.
• Self-citation drift: your own prior work cited with year off by one.

Constants

• REVIEWER_MODEL = gpt-5.5 — Used via Codex MCP. Default for cross-model review with web access.
• CONTEXT_POLICY = fresh — Each audit run uses a new reviewer thread (REVIEWER_BIAS_GUARD). Never codex-reply.
• WEB_SEARCH = required — The reviewer must perform real web/DBLP/arXiv lookups, not pattern-match from memory.
• OUTPUT = CITATION_AUDIT.md — Human-readable per-entry verdict report.
• STATE = CITATION_AUDIT.json — Machine-readable verdict ledger consumable by downstream tools.
• SOFT_ONLY = false — When true (set via — soft-only / — soft_only flag), the audit runs all three layers normally but forbids any .bib file mutation. Findings that would otherwise mutate the bib (FIX / REPLACE / REMOVE) are translated into per-occurrence sentence-rewrite proposals against the citing *.tex files. Used by /resubmit-pipeline Phase 1 to honor the user's hard "freeze the bib" constraint.
• RENDER_HTML = true — When true (default), auto-render CITATION_AUDIT.md to HTML after writing the report. Uses full Codex review gate (audit-class artifact — render-fidelity check matches the skill's cross-model audit invariant). Set false to skip, or pass — render html: false.

Workflow

Step 1: Discover bib file and section files

Locate:

• references.bib (or paper.bib / similar) under the paper directory
• All *.tex files containing \cite{...} calls (typically sec/ or sections/)

If multiple bib files exist, audit each separately.

Step 2: Extract all (cite-key, context) pairs

For each \cite{key1,key2,...} invocation in the paper:

• Record the cite key
• Record the file + line number
• Record the surrounding sentence (≥ 1 full sentence around the cite, for context check)

Output a flat list of (key, file, line, surrounding_sentence) tuples.

Also build the inverse: for each bib entry, the list of all places it is cited.

Define two protocol sets used throughout the rest of the workflow: cited_keys is the set of unique cite keys appearing in any \cite{...} invocation across the audited *.tex files (de-duplicated), and bib_keys is the set of keys parsed from the audited bib file(s). cited_keys drives Step 3 (audit only cited entries); bib_keys \ cited_keys is the uncited residual surfaced by the --uncited opt-in.

If the user passed --uncited, also compute the set difference bib_keys \ cited_keys here and stash it for use in Steps 5 and the JSON aggregation; see "Uncited Entry Detection (opt-in)" below for the protocol. The set-diff is a string operation only and does not consume reviewer budget.

Save the extracted contexts to paper/.aris/citation-audit/contexts.txt so the reviewer can read it directly. Use the paper-dir-relative path .aris/citation-audit/contexts.txt when recording the file in audited_input_hashes; do not stage under /tmp or other transient locations that the verifier cannot rehash later.

Step 3: Send each entry to fresh cross-model reviewer

For each cited bib entry — i.e., each key in cited_keys with at least one extracted citation context — invoke mcp__codex__codex (NOT codex-reply — fresh thread per entry, or batch with explicit per-entry isolation). Do not send entries in bib_keys \ cited_keys to the reviewer; those are detect-only and surface only when --uncited is explicitly enabled (see "Uncited Entry Detection" below).

mcp__codex__codex:
  model: gpt-5.5
  config: {"model_reasoning_effort": "xhigh"}
  sandbox: read-only
  prompt: |
    You are auditing a bibliographic entry. Use web/DBLP/arXiv search.

    ## Bib entry
    @article{key2024example,
      author = {...}, title = {...}, journal = {...}, year = {...}, ...
    }

    ## Where this entry is cited in the paper
    [paste extracted contexts]

    For this entry, verify:
    1. EXISTENCE: does this paper exist at the claimed arXiv ID / DOI / venue?
       Output: YES / NO / UNCERTAIN, with the verifying URL.
    2. METADATA: are author names, year, venue, title correct?
       For each, output: correct / wrong: should be ... / typo: ...
    3. CONTEXT: for each use, does the cited paper actually support the surrounding claim?
       Output per-use: SUPPORTS / WEAK / WRONG, with one-sentence reasoning.

    VERDICT: KEEP / FIX / REPLACE / REMOVE
    - KEEP: entry is clean, all uses are appropriate
    - FIX: metadata needs correction; uses are appropriate
    - REPLACE: cite is wrong-context, find a different paper that actually supports the claim
    - REMOVE: entry is hallucinated or unsupportable

    Be honest. If you cannot verify online, say UNCERTAIN; do not guess.

Save the response to .aris/traces/citation-audit/<date>_runNN/<key>.md per the review-tracing protocol.

Step 4: Aggregate verdicts

Build CITATION_AUDIT.json following the schema defined in "Submission Artifact Emission" below (single authoritative schema for this file). Per-entry ledger data goes under details.per_entry, not under a top-level entries field. The top-level verdict is a single overall value (PASS / WARN / FAIL / NOT_APPLICABLE / BLOCKED / ERROR) derived from per-entry verdicts per the decision table in "Submission Artifact Emission"; the top-level summary is a one-line human-readable string.

Concretely, details carries the per-entry ledger:

"details": {
  "total_entries": 29,
  "counts": { "KEEP": 11, "FIX": 14, "REPLACE": 3, "REMOVE": 1 },
  "per_entry": [
    {
      "key": "lu2024aiscientist",
      "verdict": "KEEP",
      "axis_failures": [],
      "uses": [
        {"file": "sections/1.intro.tex", "line": 11, "verdict": "SUPPORTS"},
        {"file": "sections/6.related.tex", "line": 8, "verdict": "SUPPORTS"}
      ]
    },
    {
      "key": "madaan2023selfrefine",
      "verdict": "FIX",
      "axis_failures": ["CONTEXT"],
      "uses": [
        {"file": "sections/2.overview.tex", "line": 42, "verdict": "WRONG",
         "note": "Self-Refine demonstrates iterative improvement, not correlated errors"},
        {"file": "sections/6.related.tex", "line": 13, "verdict": "SUPPORTS"}
      ]
    }
  ]
}

See "Submission Artifact Emission" for the full artifact (top-level fields audit_skill, verdict, reason_code, summary, audited_input_hashes, trace_path, thread_id, reviewer_model, reviewer_reasoning, generated_at, details).

Step 5: Generate human-readable report

Write CITATION_AUDIT.md:

# Citation Audit Report

**Date**: 2026-04-19
**Bib file(s)**: references.bib
**Total entries**: 29

## Summary
| Verdict | Count |
|---------|------|
| KEEP    | 11   |
| FIX     | 14   |
| REPLACE | 3    |
| REMOVE  | 1    |

## Priority Fixes (CRITICAL — apply before submission)

### REMOVE: anon2025placeholder
- Author listed as "Anonymous" — canonical record exists with real authors and full title
- Title is incomplete
- ACTION: Replace key with the canonical citekey, update authors and title

### REPLACE-CONTEXT: example2023priorwork in sec/2.overview.tex:42
- Cited to support a specific technical claim
- The cited paper actually demonstrates a different (related but distinct) phenomenon
- ACTION: Rewrite the sentence; cite the prior work for what it actually establishes

[... continues for each entry ...]

## All-Clean Entries (no action needed)

[list of KEEP keys]

When --uncited is set, append the following section after "All-Clean Entries":

## Uncited Entries (opt-in)

The following bib entries are present in the audited bib file(s) but are not referenced by any `\cite{...}` in the paper body:

- `author2010example` — suggestion: prune (uncited; no local evidence of intent)
- `someone2015othercite` — suggestion: prune (uncited; no local evidence of intent)
- `third2024todo` — suggestion: check (a `% TODO: cite third2024todo` comment was found in `sections/3.related.tex`)

This section is detect-only; it does not change the top-level verdict.

Step 6: Apply fixes (interactive)

For each FIX/REPLACE/REMOVE verdict, prompt the user:

Fix [key]?
  Change: <description of change>
  Files affected: references.bib + sec/X.tex:Y
[Apply / Skip / Defer]

If AUTO_APPLY = true, apply all FIX-level changes (metadata corrections only). REPLACE and REMOVE always require human approval — they involve content changes.

Step 7: Recompile and verify

latexmk -C && latexmk -pdf -interaction=nonstopmode main.tex

Confirm:

• No new Citation undefined warnings
• No Reference undefined warnings
• Page count unchanged or only minimally affected by metadata fixes

Uncited Entry Detection (opt-in)

Default: disabled. Existing users see no behavior change — only \cite{...} keys are audited, and bib entries with no \cite reference in the manuscript are silently ignored.

Opt-in: pass --uncited on invocation. The skill then performs a set-diff after Step 2 and reports bib entries that appear in any audited bib file(s) but are not cited anywhere in the paper. Detect-only — uncited entries are not sent to the cross-model reviewer, so there is no extra reviewer/web-lookup cost.

Why opt-in

This skill's headline output is the three-axis audit on cited entries. Surfacing uncited bib entries by default would (a) change long-form output for every existing run, and (b) noise up the verdict for users who intentionally maintain a superset bib file (e.g., shared lab bib, in-progress section reorder where the cite has been removed but the entry intentionally retained). The flag preserves zero behavior change for existing callers.

Effect when enabled

When --uncited is set:

• CITATION_AUDIT.md gains a ## Uncited Entries (opt-in) section listing the keys with a one-line suggestion each: prune (entry is dead weight; recommend deleting) or check (entry might be intentional; flag for user review). Default suggestion is prune; only emit check when there is concrete local evidence (e.g., a TODO comment in a .tex file mentioning the key, or a recently removed \cite visible in git diff). Do not infer intent from the bib key string alone.
• CITATION_AUDIT.json details gains an uncited_entries array; see "Submission Artifact Emission" below for the schema.
• The top-level verdict is unchanged: uncited entries do not upgrade or downgrade the PASS / WARN / FAIL / etc. classification. The reason_code and summary are likewise unchanged in shape; only the details.uncited_entries field appears.
• Verifier gates and downstream skills (paper-writing Phase 6, verify_paper_audits.sh) MUST NOT treat the presence of uncited_entries as a blocking signal.

When opt-in is appropriate

• Pre-submission cleanup (drop dead bib entries before sharing camera-ready ZIP).
• Shared lab bib file where the paper uses a subset and the user wants to confirm what is in scope.
• Recurring audits where the user has previously seen the uncited count and wants to track whether it changed.

Fallback when bib enumeration fails

If --uncited is enabled but full bib-key enumeration fails (e.g., malformed bib syntax that the parser cannot recover), the cited-entry audit must still proceed if at all possible. In that case:

• Do not alter the top-level verdict, reason_code, or summary.
• Emit details.uncited_entries as an empty array [].
• Add details.uncited_entries_status: "unavailable" plus a one-line note explaining why (e.g., "bib parser could not enumerate keys; cited-entry audit completed normally").
• Verifier gates and downstream skills MUST treat unavailable the same as the field being absent: not blocking.

If the bib file cannot be read well enough to audit even the cited entries, fall back to the existing BLOCKED / bib_unreadable path defined in the verdict decision table; this is the same behavior as the no-flag default.

Key Rules

• Fresh reviewer thread per audit run — never reuse prior review context
• Web access required — the reviewer must do real lookups, not memory pattern-match
• Wrong-context > metadata — a real paper used to support a wrong claim is more dangerous than a typo in author name
• REPLACE/REMOVE require human approval — never auto-modify content claims
• Always emit, never block — this skill always writes CITATION_AUDIT.json with a verdict; the decision to block finalization lives in paper-writing Phase 6 + verify_paper_audits.sh, driven by the assurance level. See "Submission Artifact Emission" below.
• Run once per submission — the audit is wall-clock expensive (web lookups for each entry); not for every save
• Uncited detection is opt-in only — never auto-enable; never block on uncited entries; existing callers must observe identical output if they do not pass --uncited
• Under --soft-only, citation-audit emits text-rewrite proposals only; bib files are never mutated regardless of finding severity. The audit semantics (existence + metadata + context) and the per-entry KEEP/FIX/REPLACE/REMOVE ledger are preserved verbatim; only the action layer is translated to per-occurrence sentence rewrites in the citing *.tex files. Refuse any downstream-proposed bib edit while --soft-only is set.

Comparison with Other Audit Skills

Together: code → result → numerical claim → cited claim. Each layer has cross-family review with no executor in the validator path.

Known Limitations

• DBLP coverage gap: very recent papers (< 2 weeks) may not yet be in DBLP. Reviewer should fall back to arXiv.
• Pre-print vs published: when both exist, reviewer should prefer the published venue (ICML 2024 over arXiv 2401.xxxxx) but flag both.
• Anthology vs OpenReview: NeurIPS/ICLR papers have OpenReview entries before official proceedings; both are valid sources.
• Multi-author truncation: bib entries with 6+ authors using and others are conventional and not flagged unless the truncation hides a co-author the user explicitly cares about.

Review Tracing

After each mcp__codex__codex reviewer call, save the trace following shared-references/review-tracing.md (Policy C — forensic; never silently skip). Use save_trace.sh (resolved per the chain in shared-references/integration-contract.md §2) or write files directly to .aris/traces/citation-audit/<date>_run<NN>/. Respect the --- trace: parameter (default: full).

Output Contract

• CITATION_AUDIT.md (human-readable report) at paper root
• CITATION_AUDIT.json (machine-readable ledger; schema below) at paper root
• .aris/traces/citation-audit/<date>_runNN/ (per-entry review traces)
• Optional: applied fixes to references.bib + sec/*.tex (with --apply flag)
• Optional: details.uncited_entries field in JSON + ## Uncited Entries (opt-in) MD section (with --uncited flag; field absent and section omitted when flag is unset)
• CITATION_AUDIT.html (when RENDER_HTML = true, default) — auto-rendered single-file HTML view via /render-html "CITATION_AUDIT.md" --json "CITATION_AUDIT.json". Full review gate. Sidecar .review.json carries render-fidelity verdict. Non-blocking: if /render-html fails (helper missing, Codex MCP unavailable, file write error), log the failure and treat the audit as complete — the JSON + MD ledger are the canonical outputs.

Optional: Soft-Only Mode (— soft-only)

Default: disabled. The audit emits the standard KEEP / FIX / REPLACE / REMOVE per-entry verdicts and a downstream caller (or the --apply path of Step 6) is free to mutate the bib.

Opt-in: pass — soft-only (also accepts — soft_only) on invocation. This mode is designed for callers — notably /resubmit-pipeline Phase 1 — that operate under a hard "freeze the bib" constraint: if a citation is wrong-context, soften the surrounding sentence; do not change, add, or remove the cite itself.

What soft-only changes

The audit semantics are unchanged: existence + metadata + context-appropriateness checks all run, the reviewer is still invoked once per cited entry, and the per-entry KEEP/FIX/REPLACE/REMOVE verdicts are still computed and emitted exactly as in default mode. Only the action layer changes — soft-only translates each base verdict into a text-rewrite proposal instead of a bib mutation.

Verdict translation table

Augmented JSON schema (under --soft-only)

When the flag is set, the standard top-level fields (audit_skill, verdict, reason_code, summary, etc.) and the existing details.per_entry ledger are emitted exactly as in default mode. In addition:

• A top-level soft_only_mode: true boolean is added.
• details gains a soft_only_actions array — one entry per audited bib key, in the same order as details.per_entry.

{
  "audit_skill": "citation-audit",
  "verdict": "...",
  "soft_only_mode": true,
  "details": {
    "soft_only_actions": [
      {
        "citekey": "smith2023example",
        "base_verdict": "REPLACE",
        "soft_action": "soften_citing_sentence",
        "occurrences": [
          {
            "file": "sec/3.method.tex",
            "line": 142,
            "current_sentence": "Smith et al. [2023] proves a generic result that...",
            "proposed_rewrite": "Smith et al. [2023] discusses a related setting; while not directly applicable, the framing motivates...",
            "rationale": "Original sentence claims smith2023 'proves' a result, but smith2023 actually only conjectures it. Softened to 'discusses ... motivates' to remove the unsupported claim."
          }
        ]
      }
    ]
  }
}

soft_action is one of keep_unchanged | keep_metadata_drift_acknowledged | soften_citing_sentence | drop_cite_in_body_only. For keep_unchanged and keep_metadata_drift_acknowledged, occurrences MAY be omitted or emitted as []. For soften_citing_sentence and drop_cite_in_body_only, occurrences MUST list one object per \cite{X} site in the body that triggered the verdict.

For drop_cite_in_body_only, the proposed_rewrite field shows the sentence with the inline \cite{X} removed (or replaced by a \cite{Y} to an alternate in-bib source). The bib_entry_action field is fixed to "leave_as_is_per_soft_only" — the bib record itself is never modified by the audit.

Augmented human-readable report

CITATION_AUDIT.md gains a new section ## Soft-Only Rewrites (— soft-only mode) listing each occurrence with the proposed sentence rewrite for human approval. Example:

## Soft-Only Rewrites (— soft-only mode)

The bib file is frozen. The following sentence rewrites are proposed in lieu of bib edits.

### `smith2023example` — base verdict REPLACE → `soften_citing_sentence`

- **File**: `sec/3.method.tex:142`
- **Current**: "Smith et al. [2023] proves a generic result that..."
- **Proposed**: "Smith et al. [2023] discusses a related setting; while not directly applicable, the framing motivates..."
- **Rationale**: Original sentence claims smith2023 "proves" a result, but smith2023 actually only conjectures it. Softened to "discusses ... motivates" to remove the unsupported claim.

The existing per-entry verdict table in the Summary block is kept but FIX/REPLACE/REMOVE rows are annotated with a 🔒 bib frozen by --soft-only badge so downstream readers see immediately why the bib was not mutated.

Hard guarantees under --soft-only

• No .bib file mutations under any circumstance. Step 6 ("Apply fixes (interactive)") is bypassed for the bib file; only *.tex rewrite proposals are produced (and still require human approval before any text edit).
• If a downstream caller — including paper-writing Phase 6 or any wrapper — proposes a bib edit while --soft-only is set, refuse it: emit a one-line refusal in the trace and continue to the next finding.
• The top-level verdict decision table is unchanged: a wrong-context cite still produces FAIL with reason_code: wrong_context. Soft-only does not silence the finding; it only constrains the action layer.
• --soft-only composes with --uncited: both flags can be set together. Uncited entries remain detect-only and are not subject to soft-only translation (there is no citing sentence to soften).

Submission Artifact Emission

This skill always writes paper/CITATION_AUDIT.json, regardless of caller or detector outcome. A paper with no .bib file or no \cite{...} usage emits verdict NOT_APPLICABLE; silent skip is forbidden. paper-writing Phase 6 and verify_paper_audits.sh both rely on this artifact existing at a predictable path.

The artifact conforms to the schema in shared-references/assurance-contract.md:

{
  "audit_skill":      "citation-audit",
  "verdict":          "PASS | WARN | FAIL | NOT_APPLICABLE | BLOCKED | ERROR",
  "reason_code":      "all_entries_keep | metadata_drift | wrong_context | hallucinated | ...",
  "summary":          "One-line human-readable verdict summary.",
  "audited_input_hashes": {
    "references.bib":             "sha256:...",
    "main.tex":                   "sha256:...",
    "sections/3.related.tex":     "sha256:..."
  },
  "trace_path":       ".aris/traces/citation-audit/<date>_run<NN>/",
  "thread_id":        "<codex mcp thread id>",
  "reviewer_model":   "gpt-5.5",
  "reviewer_reasoning": "xhigh",
  "generated_at":     "<UTC ISO-8601>",
  "details": {
    "total_entries":  <int>,                 // count of audited cited entries (= |cited_keys|), NOT the bib-file size
    "per_entry":      [ { "key": "madaan2023selfrefine",
                          "verdict": "KEEP | FIX | REPLACE | REMOVE",
                          "axis_failures": [ "CONTEXT" | "METADATA" | "EXISTENCE" ],
                          "note": "..." }, ... ]
  }
}

Optional: details.uncited_entries (only when --uncited is set)

"details": {
  ...
  "uncited_entries": [
    {"key": "<bibkey>", "suggestion": "prune" | "check", "note": "..."}
  ],
  "uncited_entries_status": "ok" | "unavailable"
}

Field semantics:

• Both fields are omitted entirely when the flag is not set. The default schema does not include either key.
• When the flag is set and the set-diff completes normally, uncited_entries_status is "ok" and uncited_entries lists the detected keys (possibly empty if every bib entry is cited).
• When the flag is set but bib-key enumeration fails (per "Fallback when bib enumeration fails" above), uncited_entries_status is "unavailable" and uncited_entries is []. Downstream consumers MUST treat "unavailable" identically to the field being absent: not blocking.
• Downstream consumers MUST treat absence of either field as the only valid default state and MUST NOT raise on missing.
• suggestion is advisory only; the verifier and paper-writing Phase 6 do not block on it.

audited_input_hashes scope

Hash the declared input set actually passed to this audit: the .bib file, main.tex, and every sections/*.tex file that supplied citation contexts. Do NOT hash extracted contexts from /tmp or other transient paths — if you need to stage extracted contexts, materialize them under paper/.aris/ so the verifier can rehash reproducibly. Do NOT hash repo-wide unions or the reviewer's self-reported opened subset.

Path convention (must match verify_paper_audits.sh): keys are paths relative to the paper directory (no paper/ prefix — the verifier already resolves relative to the paper dir; prefixing produces paper/paper/... and false-fails as STALE). Use absolute paths for any file outside the paper dir.

Verdict decision table

The --uncited flag does not appear in this table: uncited entries are advisory only and never alter the top-level verdict or reason_code. They surface exclusively through details.uncited_entries and the optional MD section.

Thread independence

Every invocation uses a fresh mcp__codex__codex thread. Never codex-reply. Do not accept prior audit outputs (PROOF_AUDIT, PAPER_CLAIM_AUDIT, EXPERIMENT_LOG) as input — the fresh thread preserves reviewer independence per shared-references/reviewer-independence.md.

This skill never blocks by itself; paper-writing Phase 6 plus the verifier decide whether the verdict blocks finalization based on the assurance level.

See Also

• /paper-claim-audit — sibling skill for numerical claim verification
• /experiment-audit — sibling skill for evaluation code integrity
• /result-to-claim — claim verdict assignment from results
• shared-references/citation-discipline.md — protocol document for citation hygiene
• shared-references/reviewer-independence.md — cross-model review constraints