ARCHIVES

Background: Contracts and online agreements contain short clauses that move money, liability, personal data and intellectual property between the parties, and a reviewer working under time pressure can easily overlook them. Language-model assistants summarise such documents fluently but seldom show the passage behind each statement, and uploading a confidential agreement to an external service may itself breach confidentiality. Objective: To design and verify a contract clause risk analyzer whose every finding is explainable, tied to an exact passage of the source document, reviewed by a person, and computed without the document leaving the reviewer's machine. Methods: ClauseGuard validates PDF, DOCX and TXT uploads by extension and content signature, extracts text with page offsets, and segments it into clauses that are exact slices of the extracted text. Each clause is matched against a versioned library of 33 hand-written rules in eight risk categories; wording in the same sentence can exclude a match or raise or lower its severity, and repeated matches are grouped and capped. A documented formula with diminishing returns converts the findings into a 0–100 review-priority score, and the reviewer confirms, dismisses or escalates each finding with a note that is stored as an auditable history. The system is built with FastAPI, PostgreSQL and Next.js and runs in three Docker containers; every tool choice and parameter is reported so that the study can be reproduced. Results: All 152 automated backend tests pass, together with a lint check, a type-checked production build and browser tests of the complete flow. On a fictional demonstration agreement the analyzer reported 25 findings across all eight categories in 56 ms, each linked to the intended clause. Conclusion: The results verify that the pipeline is functionally correct, explainable and private; they do not yet measure detection accuracy on real contracts. The paper therefore discusses the limits of pattern matching and specifies a protocol for measuring precision and recall on the CUAD and LEDGAR datasets with legal-expert validation.