/sota-auditor

name: sota-auditor
description: >
  Final authority on the truthfulness of references cited in State-of-the-Art
  reviews and academic papers. Activate this skill whenever the session
  involves auditing, correcting, or cleaning citations in a SOTA or article,
  or when responding to verdicts from `pdf-cascade` / `registry-doctor`.
  Triggers (FR + EN) : "audit SOTA", "corriger SOTA", "vérifier hallucinations",
  "purger refs", "déclarer une ref véridique ou hallucinée", "audit
  references", "verify citations", "purge hallucinations", "/paper-trail:audit-sota".
  Does NOT download (that's pdf-cascade) and does NOT write (that's sota-writer
  / paper-writer). Decides VRAI vs HALLUCINATION and applies consequences
  (correct attribution, transition to sota_cited_confirmed, or retract +
  purge from SOTA). Binary decision, no grey zone.

Skill : SOTA Auditor

Why this skill exists

Citation hallucinations in academic publications (fabricated quotes, inverted attributions, misidentified authors) are the failure mode this plugin exists to prevent. The plugin's mechanical guards (FSM, page 1 validation, cascade) reduce the surface area, but **semantic decisions remain human-driven** : « does this ref really exist ? Is this claim really in this paper ? Is this the right author ? »

`sota-auditor` is the role that makes those decisions. It is the **final authority** on the truthfulness of references cited in SOTAs / articles. It can :

• Confirm a citation (transition `page1_validated → sota_cited_confirmed`)
• Retract a fabrication (transition `* → retracted` with reason)
• Correct an attribution (modify the SOTA + the ref frontmatter)
• Rollback a technical transition by `pdf-cascade` if a homonymy was

missed

The supreme rule

**For each problematic reference, the decision is binary :**

• **VRAI** → confirm (correct attribution if needed), transition to

`sota_cited_confirmed` once the specific claim has been validated

• **HALLUCINATION** → transition to `retracted` with **mandatory

WebSearch documenting the absence** (≥ 2 independent sources confirming inexistence or error), then **purge the citation** from the SOTA(s) where it appears

**No grey zone.** No "roughly correct that we can rephrase". No "keep it just in case". An unverifiable citation is a research debt that eventually explodes.

When to activate

• The user asks for an audit of a SOTA or article bibliography
• `pipeline doctor` reports invariant violations on cited refs
• `pdf-cascade` returned `blocked_human:title_mismatch` or

`blocked_human:cascade_exhausted` on important refs

• `pipeline events --to retracted` lists newly retracted refs
• Before paper submission (pre-submission audit)
• The user mentions "vérité", "hallucination", "purger", "retraiter"
• `/paper-trail:audit-sota` command is invoked

Worker B transitions the auditor can drive

The auditor has authority over these FSM transitions. Other transitions are reserved for `pdf-cascade` (technical) or `sota-writer` (creation).

| Transition | Conditions | |---|---| | `page1_validated → sota_cited_confirmed` | Section read, verb verified, verbatim quotes if present, verdict ∈ {CONFIRMED, NUANCED}, textual evidence cited | | `* → retracted` | WebSearch mandatory (≥ 2 independent sources confirming absence/error), `retracted_reason` ∈ {hallucination, homonymy_purged, off_topic, duplicate}, purge from ALL SOTAs (verifiable by absence in markdown files) | | **Rollback** of a `pdf-cascade` technical transition | If an invariant is violated downstream (late homonymy discovery, false attribution found, off-domain missed) |

The auditor **never triggers** `∅ → candidate` (writer's job) nor the technical transitions `candidate → uid_resolved → pdf_acquired → page1_validated` (pdf-cascade's job, invokable as sub-task).

Inputs the auditor reacts to

Sources of « refs to audit » in the paper-trail workflow :

1. **`pipeline doctor` invariants** — violations on cited refs (especially I11 cited_in orphans, I12 reciprocity broken, I14 transition from terminal) 2. **`pipeline events --to retracted` or `--to page1_validated`** — refs that have moved recently and need attention 3. **`pdf-cascade` blocked outputs** — refs in `blocked_human:title_mismatch` or `blocked_human:cascade_exhausted` 4. **User-driven request** — "audit this SOTA", "audit this paper" 5. **`/paper-trail:audit-sota <path>`** command output 6. **`pipeline doctor --correlate-rtfm`** — RTFM signals that an acquired PDF has format/OCR issues

5-step method per problematic ref

Apply in order — don't skip steps.

Step 1 — Locate in the SOTA / paper

`grep -n` or `Read` to find the exact line(s) where the ref is cited. A ref is often cited multiple times — check each occurrence.

If the "ref" is cited nowhere in the actual SOTA (extraction false positive), it's a strong signal it was never a real citation.

Step 2 — Read the context

Read the complete sentence(s) citing the ref. What does the SOTA claim about this reference ? What argument depends on it ? What's the nuance (verb used, type of argument) ?

This reading is crucial : context tells you whether a candidate matches or not, and how to cleanly rephrase if needed.

Step 3 — Verify existence and attribution

Choices depending on the case :

• **Multiple candidates from `pdf-cascade`** (title-fallback returned

N matches) : compare each candidate's title to the SOTA context. Pick the one matching the sentence meaning. If none matches → step 4 without candidate.

• **Fuzzy match by `pdf-cascade`** : read the top match's title. If

semantic overlap with context is strong → confirm. Otherwise → step 4.

• **No candidate found by cascade** : external search (paper-search MCP

on multiple platforms, WebSearch on Google Scholar, arXiv) with the title, context keywords, author. Give yourself 2 searches before concluding hallucination.

• **Acronym / capitalized non-author** (e.g., "MARBLE", "HaMSE