Markdown Marty
// the markdown language model

Markdown Marty

Markdown is the model. Zero LLM at inference.

A runtime that walks markdown files — no neural net, no GPU, no embeddings, no API calls — and composes responses from cited files on disk. Hallucination is not a UX problem to mitigate. It is a category that does not exist.

Read the framework → See it on 31,100 verses →
0
LLM calls at inference
The runtime imports no model library. Enforced by static tests.
6
Architectural lie detectors
No imports. No embeddings. No subprocess. Cited files exist. Deleting changes behavior. The API serves provenance.
~30ms
Per-query latency
Cold cache. ~2ms warm. Bible-scale corpus on a laptop, no GPU.
100%
Audit traceability
Every word in every response cites the markdown file it came from. Delete the file, the answer changes.
// products in the marty family

One framework. Many domains.

If the knowledge in your domain is structured enough to write down, it is structured enough to live in markdown — and Marty can walk it.

framework · open source
Markdown Language Model
The runtime. ~700 lines of Python. Loads markdown nodes, walks them by pattern matching, composes responses by template substitution. Six tests enforce the architectural contract.
github.com/drew913s/MarkdownLanguageModel
→ Read the source
product · first reference
Bible Genius
The complete King James Bible served as a Marty graph. 31,100 verses as markdown nodes, 66 book indices, 30 topic clusters. Zero hallucination because there is no inference — only retrieval and template substitution.
github.com/drew913s/BibleGenius
→ Explore the demo
coming soon
Your domain here
Drug interaction reference. Building code lookup. Mechanical service manuals. Case law citation. Any structured-knowledge domain where hallucination is unacceptable. Marty is a category, not a project.
// fork the framework
[[ next ]]
// the architectural contract

Six tests. If any fails, the architecture is a lie.

Most "we don't use an LLM at runtime" claims rot the moment someone adds a fallback. Marty's claim is enforced by tests that run in CI — not by promises in a README.

01 No LLM imports. The runtime's Python imports must come from a stdlib + framework allowlist. Anything else fails the test.
02 No embedding or vector calls. No .encode(), cosine_similarity, faiss, chroma, or any of their cousins.
03 No subprocess. The runtime cannot shell out. Models can't sneak back in through a binary.
04 Every response cites real files. Provenance is a list of paths. Each path must exist on disk inside corpus/ or orchestration/.
05 Deleting a file changes behavior. If you can rip a node out of the corpus and the output is identical, the corpus wasn't load-bearing.
06 The HTTP API serves provenance. Every API response includes the cited corpus files alongside the answer.