2.5× the traceable reasoning of markdown at half the tokens. And Compass picks models that create DGML up to 30× cheaper than the best-scoring.
DGML vs markdown
Figures the model reasons from that point to one exact place in the document.
82–85% with DGML against 16–33% with markdown.
What the model used to answer the same questions.
6.6–7.0M tokens with DGML against 11–14M with markdown.
Creating DGML: cheapest vs best models
Model Compass finds the optimal models that turn your documents into DGML.
93.4 F1 at $0.045 per document, against 94.3 F1 at $1.33.
Traceable means each figure an answer is computed from leads to exactly one place in the document. The 30× gain comes from Model Compass choosing the models dgml runs on. How both benchmarks were run ↓
The problem
Today, documents such as PDFs are turned into markdown before they’re fed to agents. Markdown names the shape (headings, tables) but not the substance: no semantics, no typed values. An agent can cite only a heading or a string of text, and the same text often appears in many places in a document. For example, asked for the change in HII’s unamortized debt issuance costs in its 10-K, an agent reading markdown can only say the $8 million is in Note 14, near the words “unamortized debt issuance costs”, and those words appear in several places in the document. “$8 million” itself appears 10 times, and the number 8 appears 91 times.
In finance, legal, audit and compliance, a figure you can’t trace is a figure you can’t use, and a wrong one slips through.
The solution
dgml turns PDF and DOCX into DGML, semantic XML whose tags say what things are, not where they were. Each value becomes a single named concept, typed and traced. In the same example, the agent reading DGML points to
<DebtIssuanceCostAmount>
$8 million
</DebtIssuanceCostAmount>
and gives its XPath in the DGML tree:
/chunk
Path and value lead to that one element, so anyone can follow it there. Not a floating assertion: a place in the document.
Why now
Agents are moving from drafting text to producing numbers that someone has to sign off. For example, a fund’s NAV, a loan’s covenants, or a lease’s rent. In each case one party asserts a value and another must verify it independently. For example, a valuation advisor won’t stand behind a fund’s NAV unless they can trace each property’s value back to the signed lease.
For that work, an unverifiable answer is not an acceptable outcome. Cost matters as well: every token sent to an agent costs money, and markdown sends far more of them. With DGML, a document is processed once and stays fully queryable, so new questions are answered from the same file without going back to the source. DGML is an open standard with an open-source reference implementation (Apache 2.0). Try it on your own documents: github.com/dgml-io/dgml.
Verify it yourself
Both benchmarks are published in full and use public documents only.
See for yourself whether an agent’s figures point back to the document, question by question.
30 SEC 10-K filings and 233 DocFinQA questions. The same agent answers every question once reading DGML and once reading markdown. With DGML, 82–85% of the figures it reasons from point to one exact place in the document, against 16–33% with markdown, using 6.6–7.0M tokens against 11–14M.
All results are published: see the full results on GitHub. Anyone can reproduce them: rescore them offline with uv run dgr reproduce, no API key needed, or re-run the model with your own API key. See how to reproduce.
See how Model Compass compares models across forms, contracts, leases and other documents.
dgml uses one model to transcribe a document and one to label it. Model Compass ran 7 sweeps over 9 public docsets, for example forms, contracts, clinical protocols and leases, to find the most cost-efficient models for the pipeline, for each need.
| Need | Transcriber → labeler | F1 | $/doc |
|---|---|---|---|
| Volume | Haiku 5.5 → Haiku 5.5 | 79 | 0.02 |
| Balanced | Haiku 5.5 → Sonnet 5 | 83 | 0.28 |
| Best | Haiku 5.5 → GPT-5.5 | 86 | 1.42 |
Haiku 5.5 is the most cost-effective transcriber tested so far: it matches Sonnet 5, Opus 5 and GPT-5.5 (78.3–78.8) at 1/9 to 1/20 of their cost, so spend on the labeler. F1 is field values against gold, mean of 9 docsets. Up to 30× is on commercial leases: Haiku 5.5 in both roles scores 93.4 F1 at $0.045 per document, against 94.3 at $1.33 for the best-scoring pipeline.
All results are published: see the full results on GitHub. Anyone can reproduce them: re-run every sweep with uv run dgc sweep --yes, using your own API keys and dgml. See how to reproduce.