Introducing BriefGuard

The defect that changes a document is rarely a misspelled word. It is the defined term used before definition, the stale cross-reference, the duplicated clause number or the party whose name changes halfway through.

BriefGuard makes the final quality pass repeatable. It reads the document's structure, applies a versioned house profile, produces explainable findings with exact locations and proposed repairs, and gives the reviewer a controlled way to accept, dismiss or mark each issue for manual work.

The last mile is a systems problem

A long agreement may have passed through precedent banks, email, three firms and fifteen authors. Its content can be legally sound while its internal system is broken. Numbering moved but a reference did not. A schedule uses a different party abbreviation. A table quietly switched typeface. Spellcheck sees none of it; a hurried final read finds only some.

A document is not clean because nobody noticed another defect. It is clean when the quality pass is complete and evidenced.

Structure before language

BriefGuard inventories paragraphs, text runs, headings, lists, tables, definitions and cross-references before it judges a sentence. Every finding receives a stable page and structural location. That makes deterministic checks possible and keeps review grounded: the reviewer can move through the outline, see the passage in context and understand exactly which rule fired.

Six families of checks

• Defined terms — missing, unused, used before definition, inconsistent capitalisation and near matches.
• Cross-references — absent targets, wrong target types, stale heading labels and ambiguous directions.
• Numbering — sequence gaps, duplicates, hierarchy jumps and manual list numbers.
• Language — duplicate words, punctuation pairs, drafting usage and difficult passages.
• Formatting — font drift, heading mismatches, paragraph spacing and orphan headings.
• Consistency — party names, dates, currencies and modal verbs.

A desk built for judgment

The review workspace has three panes. The outline stays on the left. Findings and highlighted passages run through the centre. The inspector on the right shows severity, confidence, original text, proposed replacement and explanation. A reviewer accepts the repair, dismisses it with a reason, or marks it for manual work. High-confidence deterministic repairs can be accepted as a safe batch; language findings cannot hide inside that shortcut.

Evidence, not magic

Every pass records the profile version, enabled rules, tool and model versions, counts and quality score. Every accepted repair creates a before/after revision with actor and time. Dismissals require a note. Approval is blocked while blocker findings remain. Profiles can be cloned and versioned without rewriting completed runs. The clean-copy record and quality report therefore describe a review that actually happened.

How it is built

BriefGuard uses the System32 house stack: a Go backend with standard-library routing and SQLite, and a Next.js 16 / React 19 frontend with shadcn/ui and Tailwind v4. The demo includes five house profiles, 24 checks across all six categories, 42 documents, 46 quality runs, 1,104 findings, structural outlines, revisions, exports and activity. Tests cover seed density, individual review decisions, safe bulk fixes, run creation and blocker-gated approval.