Yes — this is the right understanding. Now convert these learnings into ClaimFlow AI’s Week 2 architecture.

Your final mental model should be:

Label Studio taught us:
data → prediction → annotation

A2I taught us:
review trigger → human loop/task → reviewer UI → human output → stored result

For ClaimFlow AI:

Document/PDF/email
→ Gemini extractedJson
→ validationJson
→ if NEEDS_REVIEW create ReviewTask
→ reviewer opens task
→ reviewer approves / edits / rejects
→ ReviewDecision stores human output
→ ReviewEvent stores audit trail

ClaimFlow AI custom human review mapping

1. Label Studio mapping

Label Studio	ClaimFlow AI
Data	`Document` row: original PDF/email
Prediction	`ExtractionRun.extractedJson`
Annotation	`ReviewDecision.correctedJson`
Prediction is read-only	`extractedJson` is immutable
Human-corrected annotation	reviewer-created `correctedJson`

Architectural decision:

Do not overwrite extraction_runs.extractedJson.
Human review creates a corrected copy.

2. A2I mapping

A2I concept	ClaimFlow AI concept
Data object	PDF/email + extractedJson + validationJson
Activation condition	`finalStatus === "NEEDS_REVIEW"`
Human loop	`ReviewTask`
Task sent to workers	task appears in `/review`
Worker UI	review detail page
Worker response	`ReviewDecision`
Output in S3	output in Postgres for now
Human loop metadata	`ReviewEvent`, `startedAt`, `completedAt`, `reviewerName`

Architectural decision:

ReviewTask is our custom human loop.
ReviewDecision is our custom human output.
ReviewEvent is our custom audit metadata.

Week 2 architecture decisions

Decision 1 — Review task creation

When validation completes and produces: