When configured for deterministic fallback, non-cancellation LLM failures produce deterministic content rather than silent success claims.
Human-in-the-loop AI
AI suggestions inside a governed research cascade
Research workflow docs describe five generated stages: gold definition, silver requirements, bronze requirements, source plan, and script plan. Human review remains the publication boundary.
Goal and questionsResearch intent enters the workflow with category and question context.
Generated stage draftsLLM-backed generation may retry repair; deterministic fallback can produce valid skeleton output.
Human stage reviewStage cards, ledgers, validators, and version comparisons keep decisions inspectable.
Safeguards to preserve
Stage payloads are validated, repair attempts are bounded, and downstream stages can be invalidated when earlier versions change.
The app surfaces final-gate context and Spark recovery actions inside the workflow detail experience.
Draft/review-needed. AI outputs are positioned as operational suggestions, not clinical recommendations, diagnoses, or automated publication approvals.