Supported composition

Build governed knowledge

Flow

This composition addresses a CRM revision structurally, resolves a JSON Pointer, checks the observed value, records the derivation, and commits a reviewable artifact. The same evidence API accepts text spans, record fields, table cells, page regions, and media time ranges.

Typed evidence to a revisioned artifact

"""Validate structured evidence, then commit a reviewable artifact revision."""

import datetime as dt

from mari_components import JsonPointer, ObjectRef, RevisionRef, ScopeRef
from mari_components.knowledge import (
    Activity,
    KnowledgeArtifact,
    LocatedEvidence,
    validate_located_evidence,
)
from mari_components.platform import InMemoryArtifactStore


def run() -> dict[str, object]:
    scope = ScopeRef(tenant="acme", space="sales")
    source = RevisionRef(
        object=ObjectRef(namespace="crm", object_id="account:42", scope=scope),
        revision="crm-v7",
    )
    material = {"plan": {"name": "enterprise", "monthly_price": 599}}
    evidence = LocatedEvidence(
        ref=source,
        locator=JsonPointer(pointer="/plan/monthly_price"),
        quote="599",
    )
    report = validate_located_evidence(
        (evidence,), resolve_material=lambda ref: material if ref == source else None
    )
    if not report.accepted:
        raise ValueError(report.issues)

    artifact = KnowledgeArtifact(
        artifact_id="fact:account-42-price",
        revision="fact-v1",
        value={"subject": "account:42", "monthly_price": 599},
        scope=scope,
        recorded_at=dt.datetime(2026, 9, 3, tzinfo=dt.UTC),
        generated_by=Activity(
            identifier="price-extractor-v1", implementation="host:model"
        ),
        evidence=(evidence,),
        derived_from=(source,),
    )
    store = InMemoryArtifactStore()
    store.commit(artifact, expected_revision=None)
    return {
        "artifact": store.get(artifact.artifact_id, scope=scope),
        "evidence": report.valid,
    }


if __name__ == "__main__":
    print(run())

Page-region and media-range evidence require a caller-supplied locator that can read the application’s chosen PDF, image, audio, or video representation. Mari validates the reference, visibility set, resolution result, and quoted material. The application assigns the acceptance policy and transaction.

Boundary	Returned value
Evidence resolution	LocatedEvidenceReport with valid rows and individual issues
Derived identity	RevisionRef containing scope, namespace, object, revision, and optional unit
Write	KnowledgeArtifact committed with compare-and-swap semantics
Adapter verification	assert_artifact_store_conforms