Context and definitions

Explicit-context decisions include only the evidence supplied or explicitly granted by your backend. Workspace files, actor memories and native conversation state are never included implicitly.

Evidence semantics

Every envelope records schema/template revisions, asserted audience, observation time, optional expiry, consistency/completeness, and opaque dependency tokens. Each item identifies its source and revision, evidence kind (observation, inference, correction, or instruction), and status. Unknown, conflicting, not-applicable and omitted values remain distinct; a known value requires an explicit value. Missing an item never establishes absence or zero.

required_records requires presence, including an explicitly unknown record. required_known requires known evidence. require_complete rejects truncated/incomplete envelopes; require_snapshot rejects a read interval. For interval reads, supply read_completed_at; do not invent a common revision. Macrofold checks these structural assertions, not the truth or completeness of the application's perception.

Use fine-grained entity tokens and query/collection tokens when membership matters. Preserve them with any application-side committed result and compare them immediately before applying effects. A single world-wide revision needlessly invalidates unrelated decisions; entity revisions alone cannot prove no new matching entity appeared. Application time is opaque data; real worker deadlines use wall time.

Immutable snapshots

Publish an envelope once, then reference its exact revision and audience:

const snapshot = await client.inferences.createContextArtifact({
  workspace_id: workspaceId,
  name: 'Case evidence, revision 7',
  context,
});
const reference = {
  artifact_id: snapshot.id,
  revision: snapshot.revision,
  audience: snapshot.audience,
};
const accepted = await client.inferences.create({
  workspace_id: workspaceId,
  definition,
  input,
  context: reference,
  model_binding,
});

References require files:read; publishing requires files:write. Content is encrypted in existing object storage and verified against its digest on read. Reference authorization runs at submission, commit and dispatch. Same-workspace audience substitution is denied. Identical actor IDs in different workspace namespaces do not share evidence. A narrower audience needs a new snapshot, not a relabelled history.

After an admitted artifact is released or permission is revoked, new disclosure stops. Revocation cannot make previously sent text unread. The typed context-reader port currently implements stored snapshots only: no arbitrary callback URLs, SQL, paths, remote registry, or automatic retrieval service.

Named definition revisions

const published = await client.inferences.createDecisionDefinition({
  workspace_id: workspaceId,
  name: 'Exception triage',
  definition,
});
const accepted = await client.inferences.create({
  workspace_id: workspaceId,
  definition: { definition_id: published.id, revision: definition.revision },
  input,
  context,
  model_binding,
});

A workspace/name/revision can be published once. A correction is a new revision. The definition pins the prompt, input/output schemas, primitive, allowed models, unknown policy, completeness requirements, ceilings, and optional bounded-tool policy. Requests may select permitted bindings and narrow ceilings; they cannot expand authority. Export the returned definition as ordinary JSON. There are no mutable aliases or visual workflow definitions in this version.

Withdrawing a definition prevents new reference resolution while retaining its immutable record and the snapshots already admitted by runs. It does not cancel an existing run; use cancellation for that purpose.

Retention and deletion

artifacts.delete(artifactId) releases independently published artifacts. inferences.deleteContextArtifact is the context-specific equivalent. A live task pin produces evidence_in_use; complete or close the task first. Physical collection follows the existing asynchronous object-storage grace policy. Deleting a diagnostic is governed by run retention, not this release operation.

Workspace purge removes published content as well as diagnostics and task evidence, preserving financial identifiers. Revoking access, expiring diagnostics, withdrawing a definition and deleting its workspace are distinct actions.

Your application must retain the minimal accepted decision/effect and required evidence under its own authorized policy if future replay needs them. An expiring diagnostic URL is not a durable application database. Retaining a decision does not authorize retaining every source conversation or restoring intentionally forgotten information.

Bring a decision-provider key

An authorized setup credential with connections:write can save an exact model account. An existing healthy OpenRouter model connection can be reused for Jev; otherwise create one:

const connection = await client.connections.create({
  name: 'OpenRouter decisions', kind: 'model', provider: 'openrouter',
  auth_method: 'api_key', secret: process.env.OPENROUTER_API_KEY!,
});
const model_binding = {
  provider: 'openrouter' as const, model: 'typesafe/jev-1.13',
  billing_mode: 'byok' as const, provider_connection_id: connection.id,
};

Include the same provider/model in the definition's allowed_models, then pass model_binding to inferences.create. Direct TypeSafe uses provider: 'typesafe', model: 'jev-1.13.0' and a TypeSafe key; the conventional adapter uses anthropic. These credentials are never interchangeable.

Save secrets only from a trusted backend or secure setup environment. A saved key is not proof of provider entitlement; the first explicitly authorized execution validates it. The invoking application's credential must belong to the connection owner. Jev remains decision-only and does not become a native harness model.