Paper Claim Audit

Paper Claim Audit: Zero-Context Evidence Verification

🔒 Do not wrap this skill in /loop, /schedule, or CronCreate. It is verdict-bearing — it judges paper-to-evidence fidelity with a deliberately zero-context fresh reviewer. Re-firing that verdict on a wall-clock timer adds no new signal (it changes only when the paper or results change). Schedule the external wait that precedes it — paper draft ready → then audit once. See shared-references/external-cadence.md.

Verify that every claim in the paper matches raw evidence for: $ARGUMENTS

Why This Exists

The executor writes experiments AND writes the paper. It "knows" what the results should be. This creates confirmation bias:

• Rounding 84.7% up to 85.3%
• Reporting best seed instead of average
• Citing metrics from a different experiment config
• Claiming "improves by 15%" when the delta is actually 12.8%

A fresh reviewer with zero prior context catches these because it has no expectations — it just compares paper text vs raw files. No external model is available here, so this "fresh reviewer" is Claude itself, deliberately re-approaching the files as a stranger would (see Step 2). This is weaker than a genuinely independent model — see the honest tradeoff note in Step 2 and auto-review-loop's "Self-Review Backend (No Second Model)" — but the zero-context file restrictions below still meaningfully reduce (not eliminate) the specific failure modes listed above.

How This Differs From Other Audit Skills

Core Principle

Zero-context, fresh reviewer. With no external model available, Claude itself is the auditor in Step 2 — but it must treat the task exactly as a fresh, zero-context reviewer would: only read what's listed below, and do not bring in anything remembered from writing the paper or running the experiments. The auditor receives (reads) ONLY:

• Paper .tex files (the claims)
• Raw result files (the evidence)

It does NOT receive:

• ❌ EXPERIMENT_LOG.md
• ❌ EXPERIMENT_TRACKER.md
• ❌ AUTO_REVIEW.md
• ❌ NARRATIVE_REPORT.md
• ❌ Any executor summary or interpretation
• ❌ Any prior audit results
• ❌ Any conversation history

This is stricter than reviewer-independence — it's zero-context evidence audit.

Workflow

Step 1: Collect Files (Executor — Claude)

Locate paper and result files WITHOUT reading or interpreting them.

Paper files (claims) — paths shown relative to the shell's working directory so you can find them with ls; when writing them into audited_input_hashes, use paths relative to the paper dir (no paper/ prefix) per the "Submission Artifact Emission" section below:

paper/main.tex                # → hash key: main.tex
paper/sections/*.tex          # → hash key: sections/*.tex
paper/tables/*.tex (if separate)   # → hash key: tables/*.tex

Result files (evidence):

results/*.json, results/*.jsonl, results/*.csv, results/*.tsv
outputs/*.json, outputs/*.csv
wandb-summary.json (if exists)
**/metrics.json, **/eval_results.json
**/config.yaml, **/args.json (experiment configs)

Exclude (no summaries, no interpretations):

EXPERIMENT_LOG.md, EXPERIMENT_TRACKER.md, AUTO_REVIEW*.md
NARRATIVE_REPORT.md, PAPER_PLAN.md, findings.md
Any .md file that is an executor-written summary

Step 2: Fresh Self-Review Audit (Claude — no second model available)

No external model is available, so this audit is performed by Claude itself, in a dedicated pass that deliberately behaves like a zero-context fresh reviewer rather than reasoning from what it remembers about the research. This is honestly weaker than a true cross-model audit — see the tradeoff note below — but the discipline of only consulting the files listed in Step 1 (and nothing else) is what does the real work here, and that discipline is followed exactly as written.

CRITICAL: Start this pass fresh. Do not carry forward reasoning, running totals, or impressions from writing the paper or running the experiments earlier in this session — re-open and re-read each file listed in Step 1 as if seeing it for the first time, exactly as a genuinely fresh reviewer would.

Adopt the following posture and protocol:

You are a paper-to-evidence auditor. Treat yourself as having ZERO prior context about this research. Read only the paper source files and raw result files listed in Step 1. Your job is to verify that every number in the paper exactly matches the raw evidence.

Paper files to read: [list .tex file paths from Step 1] Result files to read: [list .json/.csv/.yaml file paths from Step 1]

Audit Protocol

A. Extract Every Quantitative Claim

For each number, percentage, comparison, or scope statement in the paper:

• Location (section, table, caption, or inline text)
• Exact claim text
• The number or comparison being made

B. Trace Each Claim to Evidence

For each extracted claim, find the supporting raw data:

• Which result file contains this number?
• What is the EXACT value in that file?
• Match status: exact_match / rounding_ok / mismatch

C. Check These Specific Failure Modes

1. Number inflation: Paper says 85.3%, raw file says 84.7% Rule: only standard rounding to displayed precision is allowed

2. Best-seed cherry-pick: Paper says "achieves 90.2%" but that's the best of 5 seeds; mean is 87.1% Rule: check if paper specifies "average" / "best" / "median"

3. Config mismatch: Paper compares Method A vs Baseline B, but they used different hyperparameters / datasets / splits Rule: verify config files show same settings for compared methods

4. Aggregation mismatch: Paper says "average over 5 seeds" but result files show only 3 runs Rule: count actual runs vs claimed count

5. Delta error: Paper says "improves by 15%" but actual delta is (85.3 - 73.1) / 73.1 = 16.7% Rule: verify arithmetic of all relative improvements

6. Caption-table mismatch: Figure caption describes something different from what the figure/table actually shows Rule: cross-check every caption against its content

7. Scope overclaim: Paper says "consistently outperforms" but only tested on 2 datasets Rule: check if language matches actual evaluation scope

Output Format (per claim)

For each claim, report:

• claim_id: sequential number
• location: section/table/figure
• paper_text: exact quote from paper
• paper_value: the number claimed
• evidence_file: which raw file
• evidence_value: the actual number
• status: exact_match | rounding_ok | ambiguous_mapping | missing_evidence | config_mismatch | aggregation_mismatch | number_mismatch | scope_overclaim | unsupported_claim
• details: explanation if not exact_match

Overall verdict: PASS | WARN | FAIL

Honest tradeoff (see auto-review-loop's "Self-Review Backend (No Second Model)" for the full rationale): this pass has no independence guarantee. The "zero prior context" instruction reduces but cannot fully eliminate the confirmation bias described in "Why This Exists" above, because it is still the same model that wrote the paper doing the checking. Treat any claim you feel confident about without re-deriving it from the raw file as a spot to double-check rather than wave through.

Step 3: Write Report (Executor — Claude)

Parse the Step 2 audit output and write PAPER_CLAIM_AUDIT.md:

# Paper Claim Audit Report

**Date**: [today]
**Auditor**: Claude self-review, fresh zero-context pass (no external model available)
**Paper**: [paper title from tex]

## Overall Verdict: [PASS | WARN | FAIL]

## Claims Verified: [N total]
- exact_match: [count]
- rounding_ok: [count]
- ambiguous_mapping: [count]
- missing_evidence: [count]
- mismatch: [count]

## Issues Found

### [FAIL/WARN] Claim #N: [description]
- **Location**: Section X / Table Y / Figure Z
- **Paper says**: "..."
- **Evidence shows**: ...
- **Status**: [status]
- **Fix**: [specific correction needed]

## All Claims (detailed)

| # | Location | Paper Value | Evidence Value | Status |
|---|----------|-------------|---------------|--------|
| 1 | Table 2 | 85.3% | 85.28% | rounding_ok |
| 2 | Abstract | "15% improvement" | 12.8% | number_mismatch |
| ... |

Also write PAPER_CLAIM_AUDIT.json for machine consumption.

Step 4: Print Summary

📋 Paper Claim Audit Complete

  Claims verified: 24
  exact_match:     18
  rounding_ok:      3
  ambiguous:         1
  ⚠️ mismatch:      2

  Overall: ⚠️ WARN

  See PAPER_CLAIM_AUDIT.md for details.

When to Run

1. After /paper-write — first check before improvement loop
2. After /auto-paper-improvement-loop — recheck if improvement loop changed numbers
3. Before submission — final verification

Integration with Other Skills

Read by /auto-paper-improvement-loop (if exists)

if PAPER_CLAIM_AUDIT.json exists:
    read mismatched claims
    fix them as priority items in the improvement round

Advisory, Never Blocking

Same pattern as /experiment-audit:

• PASS → continue normally
• WARN → print warning, continue, flag draft as "check numbers before submission"
• FAIL → print alert, continue, but do NOT mark as submission-ready

Render HTML view (auto, when RENDER_HTML = true, default)

After writing paper/PAPER_CLAIM_AUDIT.md and paper/PAPER_CLAIM_AUDIT.json, invoke /render-html on the audit report so the user has a readable HTML view of the verdict + per-claim breakdown:

/render-html "paper/PAPER_CLAIM_AUDIT.md" --json "paper/PAPER_CLAIM_AUDIT.json"

Uses full Codex review gate (audit-class artifact — render-fidelity check matches the skill's existing zero-context cross-model audit invariant). Output lands at paper/PAPER_CLAIM_AUDIT.html with embedded source SHA256 and a .review.json sidecar carrying the render verdict.

Non-blocking: if /render-html fails (helper missing, Codex MCP unavailable, file write error), log the failure and treat the skill as complete — the JSON + MD verdict files are the canonical outputs; the HTML view is a convenience for human readers.

Skip if RENDER_HTML = false is set in the project's CLAUDE.md or passed as — render html: false.

Key Rules

• Fresh pass EVERY run. Never reuse reasoning or state from a prior audit run. Never carry context from writing the paper.
• Zero executor interpretation. Only file paths. No summaries.
• Only raw results. No EXPERIMENT_LOG, no AUTO_REVIEW, no human summaries.
• Rounding rule. Only standard rounding to displayed precision. 84.7% → 84.7% or 85% is OK. 84.7% → 85.3% is NOT OK.
• Self-review, not cross-model. No external model is available, so the reviewer is the same model as the executor -- see the "Honest tradeoff" note in Step 2. This is a real check but it does not carry the independence guarantee a genuinely different reviewer model would.

Review Tracing

After each self-review audit pass, 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/<skill>/<date>_run<NN>/. Respect the --- trace: parameter (default: full).

Submission Artifact Emission

This skill always writes paper/PAPER_CLAIM_AUDIT.json, regardless of caller or detector outcome. A detector-negative run (paper has no numeric claims) emits verdict NOT_APPLICABLE; a paper-with-numeric-claims-but-no- raw-results run emits BLOCKED. 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":      "paper-claim-audit",
  "verdict":          "PASS | WARN | FAIL | NOT_APPLICABLE | BLOCKED | ERROR",
  "reason_code":      "all_numbers_match | rounding_drift | missing_raw_results | ...",
  "summary":          "One-line human-readable verdict summary.",
  "audited_input_hashes": {
    "main.tex":                              "sha256:...",
    "sections/5.evidence.tex":               "sha256:...",
    "/abs/path/to/results/run_2026_04_19.json": "sha256:..."
  },
  "trace_path":       ".aris/traces/paper-claim-audit/<date>_run<NN>/",
  "thread_id":        "<n/a — self-review, no external MCP thread>",
  "reviewer_model":   "self (Claude, no external model available)",
  "reviewer_reasoning": "<resolved — depth actually applied by the self-review pass>",
  "generated_at":     "<UTC ISO-8601>",
  "details": {
    "total_claims":   <int>,
    "mismatches":     [ ... per-claim issue records ... ],
    "result_files":   [ ... raw files consulted ... ]
  }
}

audited_input_hashes scope

Hash the declared input set passed into this audit invocation — i.e. the exact .tex files and raw result / config files this run read — not a repo-wide union and not the reviewer's self-reported subset. If a caller passed only main.tex + a single result file, hash those two files and no others. The external verifier rehashes these entries; any mismatch flags STALE.

Path convention (must match what verify_paper_audits.sh expects): keys are paths relative to the paper directory (the arg passed to the verifier) for in-paper files — so main.tex, not paper/main.tex — and absolute paths for out-of-paper files such as external results/ dirs. The verifier resolves relative entries via os.path.join(paper_dir, key); prefixing with paper/ produces paper/paper/main.tex and false-fails as STALE.

Verdict decision table

Fresh-context discipline

Every invocation is a fresh self-review pass with no external model. Do not accept prior audit outputs (PROOF_AUDIT, CITATION_AUDIT, EXPERIMENT_LOG, AUTO_REVIEW summaries) as input to this audit — the zero-context restriction in Step 2 is what this skill has instead of true reviewer independence per shared-references/reviewer-independence.md; carrying prior audit context in would remove even that.

Human-readable sibling

paper/PAPER_CLAIM_AUDIT.md is written alongside the JSON for readers. The JSON is authoritative for verify_paper_audits.sh; the Markdown is for humans. The parent skill (paper-writing Phase 6) plus the verifier decide whether the verdict blocks finalization — this skill itself never blocks; it only emits.