Skip to main content

Workflow Run Record

Each accepted FactVerse AI Agent workflow run should leave a clear record. The record helps operators review decisions, engineers improve workflows, and customers trace how an answer or action was produced.

Prerequisites

Create a run record when:

  • a workflow uses operational source data, documents, simulation results, or work records;
  • an Agent output may influence inspection, maintenance, facility operations, simulation planning, or operator guidance;
  • a human reviewer accepts, revises, or rejects the output;
  • source quality, missing evidence, or scope limitations affect the answer;
  • the result should feed future model tuning, knowledge updates, or operating playbooks.

Before the run starts, prepare the workflow owner, review owner, endpoint, scopes, target boundary, expected output type, and record location.

Record structure

SectionRequired content
RequestUser request, workflow type, boundary, time window, and expected output.
AccessEndpoint, visible tool set, scopes used, and client identity.
EvidenceSource systems, records, timestamps, documents, scene versions, and tool outputs used.
AnalysisFacts, assumptions, limitations, missing data, and generated recommendation.
ReviewReviewer, decision, required changes, and approval state.
ActionDraft or accepted work order, inspection task, scenario package, or validation record.
FeedbackField result, correction, false positive, repaired data, or follow-up note.

Expected output

A complete run record should let another reviewer answer four questions without rerunning the workflow:

QuestionEvidence in the record
What was asked?Request, boundary, time window, user role, and expected output.
What evidence was used?Source systems, record IDs, timestamps, scene versions, documents, and tool outputs.
What did the Agent infer?Confirmed facts, assumptions, limitations, missing data, and generated recommendation.
What did the reviewer decide?Review state, reviewer, approved action, rejection reason, and feedback.

Template

Workflow:
Boundary:
Time window:
Requested by:
Review owner:

Endpoint and scopes:
Runtime-visible tools:

Source evidence:
- Source:
Record:
Timestamp:
Notes:

Generated output:
- Confirmed facts:
- Assumptions:
- Missing data:
- Recommendation:

Review decision:
- Accepted / revised / rejected:
- Reviewer:
- Reason:
- Follow-up:

Final record:
- Work order / inspection / scenario / validation ID:
- Feedback captured:
- Next review date:

Workflow-specific fields

WorkflowAdd these fields
Facility operationsSite, area, asset ID, alarm IDs, inspection IDs, work-order IDs, affected nearby equipment, and operator notes.
Predictive maintenanceEquipment ID, component, operating mode, signal window, health or anomaly output, maintenance history, and engineer feedback.
Physical AIScene version, model asset versions, component geometry, simulation backend, asset-readiness issues, assumptions, validation notes, and reuse target.

Review states

StateMeaning
DraftAgent output is available but has not been reviewed.
Needs data correctionOutput is blocked by missing, stale, or mismatched source data.
Needs engineering reviewOutput depends on assumptions, compute results, or domain interpretation.
Accepted with limitsReviewer accepts the output for a defined boundary and records limitations.
Accepted for actionReviewer approves the draft action or handoff record.
RejectedReviewer rejects the output and records the reason.

Evidence rules

  • Keep source timestamps visible.
  • Keep generated recommendations separate from confirmed facts.
  • Record the endpoint and scopes used by the client.
  • Keep draft actions separate from approved actions.
  • Capture reviewer corrections so future runs can improve.
  • Reuse Physical AI results only when scene version, asset version, runtime parameters, and validation notes are recorded.

Example record summary

FieldExample
WorkflowPredictive maintenance review
BoundarySite A, compressed air system, compressor C-03, last 14 days
EvidenceSignal trend, anomaly output, inspection record, work-order history, and maintenance note
OutputBearing temperature trend increased, vibration signal has two missing intervals, inspection recommended before planned shutdown
ReviewMaintenance engineer accepted inspection recommendation and marked missing vibration intervals for data correction
FeedbackInspection found lubrication issue; work order closed with parts and field notes attached