--- name: cn-ib-cross-document-check description: Reconcile facts, figures, dates, names, definitions, and disclosures across multiple Chinese investment-banking documents. Use when comparing 招股说明书、发行保荐书、上市保荐书、审计报告、法律意见书、募集资金运用文件、问询回复、再融资申报材料 or different document versions for 跨文件一致性、数据勾稽、口径核对、版本差异 or contradiction detection. Use cn-ib-filing-proofreader for single-document language and formatting review. --- # CN IB Cross-Document Check Compare multiple filing documents without flattening differences in period, entity, scope, unit, or accounting definition. ## Non-negotiable rules 1. Compare like with like. Match entity, metric definition, period, scope, currency, unit, and pre/post-transaction basis before declaring a conflict. 2. Preserve provenance for every extracted fact: document, version, page/location, source text, and extraction method. 3. Treat rounding differences separately from substantive inconsistencies. 4. Do not decide which document is correct merely because one is newer. Identify the likely source of truth and request confirmation. 5. Distinguish a missing disclosure from a contradictory disclosure. 6. Escalate legal, accounting, and regulatory judgments to the responsible professional. ## Workflow ### 1. Build the document manifest Read `references/document-matrix.md`. Record for every file: - document type; - filename and version; - document date; - reporting period; - responsible owner; - expected source-of-truth fields; - whether the document is draft, signed, filed, feedback response, or registration version. Never compare files until their versions are identified. ### 2. Define the comparison dictionary Start with the high-risk fields relevant to the transaction: - issuer and subsidiary names; - incorporation and historical dates; - share capital, ownership, and per-share figures; - revenue, profit, cash flow, assets, liabilities, and margins; - customers, suppliers, related parties, and contracts; - market share, ranking, capacity, production, and sales volume; - fundraising project investment, proceeds, implementation entity, and schedule; - risk factors and material event descriptions; - sponsor, counsel, accountant, and signatory details. Define each metric before extraction. Read `references/comparison-schema.md`. ### 3. Extract an evidence ledger Create one row per claim using: ```text document, version, entity, metric, definition, period, scope, value, unit, location, source_text, extraction_method, reviewer ``` Do not normalize away the original text. Store both the original value/unit and the normalized value/unit. ### 4. Run deterministic comparison For structured numeric claims, use: ```bash python3 scripts/compare_numeric_claims.py claims.csv --output conflicts.csv ``` The script normalizes common Chinese currency and share units and applies a configurable tolerance. It does not determine the authoritative value. ### 5. Review contextual conflicts Read `references/conflict-rules.md`. For every script result or manually observed difference, decide whether it is: - a true contradiction; - a period mismatch; - an entity or consolidation-scope mismatch; - a pre/post-transaction mismatch; - a currency or unit mismatch; - a rounding difference; - a definition mismatch; - an intentional update; - a missing disclosure; - insufficient evidence. Quote the minimum relevant source text from each document. ### 6. Identify the likely source of truth Use the document matrix and transaction context. Examples: - audited financial figures usually trace to the latest signed audit report and underlying ledger; - share capital traces to current corporate records and the applicable transaction assumption; - fundraising amounts trace to the approved project and filing version; - legal entity names trace to valid registration documents; - market data traces to the cited external source and specified date. Phrase the conclusion as `建议以……为准,并由……确认`, not as an unsupported final determination. ### 7. Produce the reconciliation report Use this structure: | ID | Severity | Topic | Document A | Document B | Difference | Classification | Likely source of truth | Required action | Owner | |---|---|---|---|---|---|---|---|---|---| Also provide: - counts by severity and classification; - unresolved P0/P1 items; - documents or appendices not received; - fields excluded from comparison and why; - a closure column for corrected version and reviewer evidence. ### 8. Apply approved corrections When asked to correct existing DOCX files: - preserve each original and edit the identified authoritative version only; - apply native Word tracked insertions and deletions by default; - change the smallest necessary text, number, unit, or table cell; - preserve existing revisions and comments unless instructed otherwise; - do not simulate redlines with color or strikethrough; - record the corrected version and exact location in the reconciliation report; - structurally verify the revision markup, re-render changed pages, and rerun the comparison. Use a transparent revision author such as `Codex` when none is specified. Produce a clean accepted copy only when explicitly requested. ## Completion gate Do not call the review complete until: - every document has a recorded version; - all P0/P1 items have an owner and action; - every numerical conflict retains source locations; - rounding and definition differences are distinguished; - corrections are rechecked in the revised files; - visual or scanned content excluded from extraction is disclosed. Use temporary storage for extraction and comparison files. Keep only user-requested final artifacts in the deal directory.