Hi MDXG team — sharing a small piece of related work in case it's useful to the community. No specific ask.
We tested whether the dual-audience premise from your README — "markdown is the primary interface between humans and AI agents […] markdown is already ideal for agents: low token cost, simple to generate, trivial to parse. The missing piece is the human side" — holds empirically on the agent side.
Setup: 558 atomic claims projected as <entity_id>/<predicate>.md files (one fact per file, literal value as body). One model (Sonnet 4.6) in Claude Code (--print), three retrievers in isolation, 28 auto-verifiable questions:
- R1 — SQL exact on the source Postgres
- R2 — MCP grep tool (JSON inline)
- R3 —
grep/ls/cat/findon the markdown projection
Findings: R3 (file-grep on the projection) wins on precision (71% exact-match vs 68% / 57%), latency (34s vs 22s / 99s), and runtime robustness (0 errors vs 0 / 5 for SQL). The agent side of the dual-audience model just works with stock Unix tools — on the same files you spec for human consumption.
The paper's §5.7 "Related work — MDXG" credits the spec explicitly and quotes the motivation line above. Any .md produced by our projection script is MDXG-conformant by construction; a viewer like @mdxg/vscode opened on the same tree would give operators the polished navigation MDXG specifies, while agents grep in parallel. One substrate, two consumers.
- Paper (HTML, anonymized canary): https://guillaume-hestia-projekt.github.io/svl/research/2026-05-retriever-bench/
- Source paper + reproducible bench: https://github.com/guillaume-hestia-projekt/svl/tree/main/research/2026-05-retriever-bench
- Methodology builds on Krishna et al. (arXiv 2605.15184) extended to a small G1 fine atomic-claim regime not covered by that paper.
Figured the empirical agent-side confirmation might interest the MDXG community. Happy to close if off-topic for the issue tracker — Discussions aren't enabled so this was the only available channel.
Activity
0No activity yet.