UN DocFormat Agent

UN and UN-affiliated organizations produce thousands of formal documents a year under one of the world's strictest editorial standards: the DGACM Instructions and UN Editorial Manual. Word limits, consecutive paragraph numbering, UNTERM country names ("Viet Nam", not "Vietnam"), Oxford spelling ("programme", "-ize")— today all checked by hand. UN DocFormat Agent automates the entire pipeline: 1. Rules engine (15+ checks, pure Python) — validates a draft against the citable UN standard and reports every violation with rule ID, location, and exact fix. No hallucinated rules: docs/UN_RULES.md maps every check to its DGACM source. 2. Deterministic autofix — mechanical violations (spelling, terminology, currency, quotation style, numbers, sub-listing markers, renumbering) are fixed without any model, each recorded in a change log. 3. AI rewrite on AMD, with a regression guard — violations needing editorial judgement go to MiniMax M3 running on AMD Instinct GPUs via the Fireworks AI API. The model is constrained to fix only the flagged issues, quoted material and technical numbers are protected, and the rules engine re-verifies the result(The client is model-agnostic — one env var swaps in any Fireworks-hosted model.) 4. UN-compliant Word output — 10-pt Times New Roman, US letter, 1" margins, bold headings, correct sub-listing indentation — plus a SEPARATE change-log .docx for reviewers, always included. The editor mirrors every result with undo/redo. Fully containerized (Docker), 67-test suite, MIT-licensed, live demo deployed. The architecture extends to other UN document types (resolutions, briefing notes) by adding rules, not code — and to any organization with a house style: the same engine could enforce ISO, EU, or corporate editorial standards. Note for judges: the live demo is on a free-tier host — first load may take ~1 minute if the instance is cold.