control.human_review: Pause the run for selective field confirmation or correction, then continue with the reviewed data.
Use this step after extraction or transformation when a human must confirm or
correct specific fields before the run continues. Place it as a top-level
sequential step only — nested placement inside other control containers is
rejected at publish time.
metadataFrom: ai.extract is set and data references the whole extract
output, the step strips the reserved _grounding key, derives categorical
confidence (low, medium, high) per field from nested dotted-path
grounding, and reuses the extract schema when schema is omitted.
When data is exactly {{ steps.<extract-step>.output }}, Eigenpal also
snapshots any lineage and ParsedDocument sidecars the extract step produced.
Studio loads them for the review workspace: source PDFs/images beside the
structured output, spatial citation highlights on supported fields, and bbox
overlays when a canonical ParsedDocument is available. Missing or unreadable
sidecars are skipped — review still works from _grounding alone. Review
decisions are not yet written back into lineage.
The human review API includes optional lineage and parsedDocument on task
detail when sidecars resolve; they are API-only context, not workflow output.
Selection modes: confidence reviews fields below a threshold;
explicit reviews paths matched by include/exclude; all reviews every scalar
leaf. Thresholds accept numeric 0–1 or categorical low|medium|high.
Per-path overrides live under selection.fields with review: always|never|skip and optional per-field threshold.
Selectors: pointers address scalar leaves, not containers. For an array of
strings named reasons, use /reasons/* (all items) or /reasons/0 (one item),
not /reasons. For nested objects, select a leaf such as /client/name.
include, exclude and selection.fields follow the same rule. Validation
rejects selectors that a known review schema proves cannot reach a scalar,
including schemas inferred through metadataFrom: ai.extract. Dynamic or
unresolved schemas are checked at runtime; even a valid array selector can
match no fields when that array is empty.
Evaluations: evaluation runs always simulate review and never pause. Assert
routing and simulated edits with review fixtures in dataset meta.json — see
Build a dataset.
Configuration
Configuration goes inside the step’swith: block.
string | record<string, unknown> | array<unknown>
required
Structured data or a template expression resolving to an object or array
record<string, unknown>
Optional JSON Schema for validating edits and exposing downstream autocomplete. When omitted with metadataFrom: ai.extract and a direct whole-output data expression, the extract step schema is reused.
"ai.extract"
Producer adapter.
ai.extract derives field confidence from _grounding and strips that reserved key from review data. Grounding is not stripped unless this is set.record<string, object>
default:"{}"
RFC 6901 pointer map of producer facts (labels, confidence, display). Selection policy belongs under selection.fields; fieldMetadata.review is a legacy alias.
object
required
Field selection policy resolved once when the task is created
object
default:"{\"includeRunInputs\":true,\"attachments\":[]}"
Run input files and optional current-run attachments for reviewers
string
Free-form guidance shown above the review workspace
Output
Returnsunknown. The complete reviewed data in the same object or array shape as the input.