Documents in, documents out

Much enterprise data arrives as prose: reports, forms, emails, contracts. GND treats a document as quoted data. An agent extracts facts from it into a population that the model checks, and the agent never follows instructions written inside the document. The reverse direction, rendering a document from known facts, gives you test material whose ground truth you control.

Read a document into facts

Use gnd-extract with model.orm.yaml and artifacts/ to extract report.md into
extracted.population.json. Preserve the document's facts, including facts that
violate the model. Write the digest-bound extraction sidecar, then run the verifier
and explain its findings.

1. gnd extract model.orm.yaml --out artifacts writes the extraction vocabulary: a shape-only JSON Schema, a constraint-blind verbalization and a header template. It runs offline and never reads the document. With --guidelines it also writes a constraint-blind guide to what each fact type and role means.
2. The agent writes extracted.population.json and a sidecar that binds the model, the document and the vocabulary by digest.
3. gnd verify model.orm.yaml extracted.population.json --artifacts artifacts checks the handoff first (a missing or stale sidecar is refused at band 3), then the facts.

The vocabulary is deliberately constraint-blind. If the document describes an invalid state, the extraction must preserve it so that the checker can reveal it. The agent is not told how to make the facts pass.

To pin important values to the text, an anchored constraint checks that each value has a witness span in the bound source document (gnd check … --anchors … --document report.md). The sidecar format is in the anchor sidecar reference; to combine several anchored documents into one population, see composition.

Write a document from facts

Use gnd-render-doc to render a status report from seed.population.json for
model.orm.yaml. Plant all three distractor kinds, write the manifest, and run doccheck
until it passes.

The skill reads the population and its verbalization (never the model file), writes document.md, and plants three kinds of distractors so that an extractor which grabs everything in sight is caught:

gnd doccheck model.orm.yaml document.manifest.json --artifacts artifacts verifies the rendering obligations: every applicable fact is mentioned, and every recorded distractor really appears in the text and is really absent from the source facts.

Measure the round trip

Render, extract and compare: gnd fidelity measures, per fact type, how faithfully extraction recovered the source population across committed round-trip cases. It keeps three questions separate: whether the rendering was complete, whether the extracted facts satisfy the model, and whether the extraction was faithful.

The documents guide, the extraction contract and the rendering contract define the exact sidecars, manifests and failure semantics.