Building with AI
UW Markdown is an open standard, not a hosted AI service or a single application. An AI agent works with the same .uwx.md record that calculation engines, underwriting systems, and editors use, and loads it directly. No custom connector or MCP server is required.
The standard draws one firm line. AI can extract, classify, and explain. Deterministic code calculates and verifies the financial results. An agent reads the rent roll and T-12, writes facts and narrative into the record, and leaves every NOI, DSCR, and IRR to the calculation engine.
- 01 · SourcesDeal documents
Rent roll · T-12 · lease abstracts · loan terms · market assumptions
- 02 · Analyst or AIExtract and explain
Pull facts from documents, classify them, flag gaps, draft the credit narrative
- 03 · Record
.uwx.mdFacts, assumptions, narrative, and provenance in one file both sides read
- 04 · EngineValidate and calculate
Check structure and rules, then compute the metrics a registered calculation pack or verifier defines
- 05 · ResultsRecomputed figures
DSCR · debt yield · LTV · cap rate · verified cash-flow returns, with an optional verification receipt
- 06 · ConsumersEvery downstream system
Excel · credit memos · underwriting apps · APIs · portfolio and data systems
AI is optional. A person, a parser, or an application can write the same record. Whoever writes it, financial results come from deterministic code, not from the model's output.
Receipts prove derivation, not truth. A verified receipt shows the stated results follow from the record under a named pack. It does not show that the rent roll or T-12 behind it is accurate.
Rules for agents
- Preserve the Markdown narrative and labeled JSON blocks.
- Extract source facts, classify information, identify gaps, and draft narrative.
- Never calculate financial results. NOI, DSCR, LTV, debt yield, cap rate, IRR, NPV, and DCF come from deterministic code or formulas.
- When editing, preserve bytes outside the requested edit region.
- Preserve provenance. Append or supersede; never silently destroy history. The host, not the agent, writes each block's
_meta. - Rates are fractions (
0.055means5.5%). - Treat the format specification, protocol, metric dictionary, schemas, and conformance fixtures as the contract.
Agent execution belongs to the producer. The reference toolkit provides non-normative context helpers and generic writers (RFC 0078). These rules come from the non-normative agent writers guide, and the reference toolkit's generic writers apply them. Rule 3 binds every tool, and the standard keeps its provenance rules (append or supersede, host-owned _meta). The reference toolkit runs no agent: your host runs its own model and writes the result through writeAgentBlock or uwmd edit. See tools and packages.
Best context order
llms.txtfor discovery.- Quickstart and the first-record tutorial for orientation.
- Format specification for document syntax and semantics.
- Protocol specification for Reader and Verifier requirements and optional verification profiles; the starter guide describes the writer conventions and context profiles.
- Schemas and conformance fixtures for implementation testing.
llms-full.txtfor the expanded link map.
Optional HTTP and MCP profiles
The HTTP binding, OpenAPI contract, and MCP binding describe interoperable server shapes. They are optional profiles. This site does not expose live deal resources or mutation tools.
The MCP profile defines uwmd.get_document, uwmd.validate, uwmd.convert, uwmd.apply_edit, and uwmd.list_representations.
Ready-to-use instruction files
- Codex-compatible skill
- Claude instructions
- ChatGPT project instructions
- Gemini instructions
- Platform-neutral guide
Canonical project
- Repository: https://github.com/UWMD-OSP/UW-Markdown
- License: MIT
- Authoring: UW Format 3.0 (
.uwx.mdfor structured records); readers refuse any otheruw_version. - Compatibility: see the current version matrix for protocol and package versions.