Anti-Autoresearch

End-to-end substantive-integrity forensic sweep of a research paper (especially autoresearch / AI-Scientist-style output). Orchestrates the whole pipeline: ingest (arxiv-id | pdf | dir → working dir + pdftotext for L0) → /evidence-ledger (artifact manifest + observability level L0/L1/L2 + span-anchored claims.json) → fan out the integrity auditor skills (consistency, citation, baseline, experiment, presentation, proof-derivation, eval-design — each reads the ledger, emits span-anchored findings) + the zero-verdict-weight AIS writing-style track → advisory memos (/adversarial-case-builder + /novelty-duplication-advisory, no verdict weight) → deterministic tools/adjudicate_findings.py (--ledger REQUIRED) → reviewer-ready Integrity Forensics Report. Cross-model (fresh codex per dimension) and reviewer≠adjudicator: the model proposes findings, the deterministic adjudicator decides the verdict. Observability-aware, detect-only, never an opaque AI-text classifier (a separate zero-weight AIS section lists AI writing

Flagship intra-paper self-consistency forensics: does the paper contradict ITSELF across abstract/intro/tables/body/appendix, and does the method DESCRIBED match the method EVALUATED? Needs no external ground truth — works PDF-only (L0). Runs a deterministic arithmetic pass + a fresh cross-model semantic pass, every finding span-anchored to the evidence ledger (claims.json), reviewer≠adjudicator. Emits consistency-audit.findings.json; NEVER computes the verdict. Triggers: "consistency audit", "check the paper against itself", "self-consistency", "内部自洽".

Audit whether a paper's baseline comparisons are COMPLETE, FAIR, and SIGNIFICANT: a required recent SOTA baseline is missing while 'best/SOTA' is claimed (HP-MISSING-BASELINE); a baseline is undertuned / given less compute-tuning-data, run at a mismatched config, or the equal-budget ablation-as-baseline is absent (HP-WEAK-BASELINE); 'outperforms' is asserted over overlapping error bars or with no variance/seeds (HP-SIG-OVERLAP); and a cross-row 'improves over baseline by X%' is arithmetically wrong (HP-DELTA-ERROR, cross-row form only). A versioned per-domain baseline profile + a live leaderboard/recency search are assembled by the EXECUTOR as structured facts; a fresh cross-model reviewer (gpt-5.6-sol xhigh, read-only, fresh thread per dimension) PROPOSES findings, each span-anchored to a ledger claim_id; tools/adjudicate_findings.py DECIDES the verdict. Works at L0 (stated comparisons) and deepens at L2 (configs/result files). A completeness question it cannot settle internally becomes needs_external_check,

Citation-integrity forensics: is every reference real, correctly attributed, and used in a context the cited work actually supports? Catches hallucinated references (no paper at the claimed arXiv id/DOI/venue, fabricated authors/year), metadata drift (wrong year/venue/version), and wrong-context citations (a real paper cited for a claim it never makes — or argues against). A hot zone for machine-generated papers. Decidable at L0 (text + canonical sources). Span-anchored to the evidence ledger (claims.json); the executor gathers canonical facts (DBLP / arXiv / DOI), then one FRESH cross-model thread per cited key proposes findings; reviewer != adjudicator. Emits citation-forensics.findings.json; NEVER computes the verdict. Triggers: "citation forensics", "check the references", "hallucinated citations", "wrong-context citation", "verify references", "引用核对".

Audit whether a paper's EVALUATION DESIGN actually measures what it claims and whether its reporting is complete — the validity layer family D (experiment-forensics) cannot reach. Three patterns: train/test leakage means the reported score may not measure generalization (HP-EVAL-LEAKAGE — adopts the Kapoor & Narayanan 8-type / 3-category leakage taxonomy; the illegitimate-proxy / sampling-bias / pretraining-contamination subtypes hand off as needs_external_check, naming but NEVER running Oren-2023 exchangeability / Shi-2023 Min-K% / Golchin-2023 Time-Travel / BIG-bench canary); a load-bearing LLM judge is conflicted (same model/family as a compared system) or unvalidated (no human-agreement, no bias control) (HP-JUDGE-VALIDITY); a declared condition/metric is dropped or switched to favor the method, or 'best' is chosen with no held-out set (HP-SELECTIVE-REPORTING). Verdict-bearing at L0/L1 from the DESCRIBED protocol — NOT repo-gated like experiment-forensics; L2 only CONFIRMS against split/preprocessing/resu

Audit experiment integrity against the evidence ledger. At L2 (repo + result files present) a fresh cross-model reviewer reads the eval code line-by-line for fake/derived ground truth, score self-normalization, phantom results (a paper number with no backing file/key), dead/uncalled metric code, verified-scope inflation, method-described ≠ method-evaluated drift, synthesized-looking results, placeholder/fake data still wired into a released result, code-output ≠ reported-number mismatch, and missing reproducibility artifacts (an empirical/agent/LLM paper shipping neither code nor the prompts/configs its results need) — every finding span-anchored to a ledger claim_id. At L0/L1 (PDF / source only) the same patterns are surfaced as info-level 'could-not-verify' signals where the ledger gives an anchor (observability_level_required:2) — NEVER a fraud verdict from a PDF. The reviewer PROPOSES findings; tools/adjudicate_findings.py computes the verdict. Detect-only. Triggers: "experiment forensics", "audit the res

Family-G proof & derivation integrity forensics: does a THIRD PARTY's written proof/derivation actually establish its theorem, or does it skip an obligation, assume its own conclusion, take an invalid step, drift a symbol's meaning, or smuggle an unstated assumption? Decides from the WRITTEN proof/derivation — verdict-bearing at L1 (the LaTeX source; PDF-extracted math is unreliable, so an L0 PDF-only run surfaces info only) — never asserts 'fabricated', only that the step shown does not hold. A fresh cross-model reviewer reads the theorem/proof + an extraction-only obligation scaffold and proposes per-obligation findings, each span-anchored to the evidence ledger (claims.json); reviewer≠adjudicator. Emits proof-derivation-forensics.findings.json; NEVER computes the verdict. dimension=proof, can be critical. Triggers: "proof forensics", "check this proof", "derivation integrity", "audit the math", "证明审计", "推导有没有漏洞".

Synthesize the single strongest EVIDENCE-BOUND reviewer case to reject a paper, built ONLY from the evidence ledger (claims.json) + the other auditors' confirmed findings — never free-floating LLM critique. Two fresh cross-model codex threads: an attack writes the ~200-word rejection paragraph (every accusation tagged to an existing claim_id/finding_id), a defense decomposes it and rules each point against the anchored evidence. MEMO-ONLY: emits adversarial-case-builder.memo.md (fed to the adjudicator via --memo) and carries NO verdict weight — tools/adjudicate_findings.py lists it in MEMO_ONLY_SKILLS and caps it at info. Honest-null allowed (the paper may survive). Run LAST. Detect-only. Adapted from ARIS kill-argument. Triggers: "adversarial case", "strongest objection", "rejection memo", "kill argument", "最强拒稿点".

Transparent, itemized impressions of AI-generated WRITING STYLE — the repo's ONLY non-integrity track. Two passes: a deterministic defensive-hedge density screen (tools/check_ai_style.py, AIS-DEFENSIVE-HEDGE) plus a fresh cross-model GROSS-cases-only semantic pass over the 13 AIS-* style tells (broken narrative arc, LLM phrase tics, jargon-stuffing, invented codenames, clause/formula walls, gratuitous pseudocode, bullet overuse, bold-module spam, restatement loops, focus drift, single-style figures, appendix dumping). Every finding is named, LOCATED, span-anchored to the evidence ledger (claims.json), carries not_integrity_finding:true + false_positive_risk:high + an fp_case, and gets ZERO verdict weight: the adjudicator forces it to info, excludes it from overall_verdict, and renders it in a SEPARATE report section. NOT an AI-text classifier — no scores, no "this is AI-written", no authorship probability; a paper can be CLEAN_GIVEN_EVIDENCE and still list many. Emits ai-style-impressions.findings.json; compu

Build the deterministic evidence ledger (artifact_manifest.json + claims.json) that every other Anti-Autoresearch auditor reads. One pass inventories artifacts, derives the observability level (L0 PDF-only / L1 +LaTeX / L2 +repo+results) by fixed rule, and extracts span-anchored, hashed, checkable claims (numbers, comparisons, scope, method, baselines, citations, captions, table cells) into claims.json. An OPTIONAL additive cross-model pass ADDS span-anchored semantic claims — method, theorem statements with their assumptions, definitions, proof/derivation steps and equations, scope, baselines, conclusions, the motivation span, and reproducibility-artifact references (the proof, derivation, and structure anchors the family B/D/G auditors need) — it never invents a number, emits a finding, or computes a verdict. Run FIRST, before any audit skill. Triggers: "build the ledger", "extract claims", "prep for integrity audit", "evidence ledger", "建证据账本".

MEMO-ONLY prior-work overlap advisory: surfaces the two ADVISORY taxonomy signals neither a tool nor a model can decide from the paper alone — ADV-TRIVIAL-COMBINATION (standard A+B+C / 缝合 stapling) and ADV-DUPLICATE-PUBLICATION (repackaged / duplicate submission). The executor RETRIEVES candidate prior work (DBLP fuzzy-title + boolean · WebSearch · WebFetch) from the paper's own title + contribution spans in the evidence ledger; TWO fresh cross-model codex reviewers (one per axis) LAY OUT the overlap side-by-side against each anchored contribution claim. It NEVER rules 'trivial' or 'duplicate' (that is a human judgment) and absence of a match is NOT evidence of originality. Emits novelty-duplication-advisory.memo.md + an info-only findings mirror; carries NO verdict weight — tools/adjudicate_findings.py lists it in MEMO_ONLY_SKILLS and caps it at info. Detect-only. Adapted from ARIS novelty-check, reframed from 'is MY idea novel' to 'here is the overlap a reviewer should weigh'. Triggers: "novelty advisory",

Checkable-ish surface presentation signals a reviewer notices first — duplicate/near-identical tables, leftover pipeline/template strings, too-few or LLM-looking figures, and page-padding. AUXILIARY ONLY and weak by design: a deterministic pass (tools/check_presentation.py — dup-table + pipeline-artifact) plus a fresh cross-model GROSS-cases-only semantic pass (thin-float, LLM-figure, page-padding), every above-info finding span-anchored to the evidence ledger (claims.json). The adjudicator CAPS everything at minor (SURFACE_ONLY_SKILLS + SURFACE_PATTERNS) — these contribute at most SOFT_FLAGS, never a HARD verdict — default false_positive_risk:high. NOTE: the pure AI writing-STYLE impressions (AI-flavor prose, defensive 'not-X-but-Y' hedging, narrative-arc, jargon-stuffing, invented codenames) MOVED to the zero-verdict-weight AIS track — for those use skills/ai-style-impressions, NOT this. Emits presentation-signals.findings.json; NEVER computes the verdict. Triggers: "presentation signals", "surface check",