The UI/orchestrator works as follows:

• On each run, Autonima computes deterministic hashes from the normalized config/schema, including per-stage hashes for search, screening, retrieval, parsing, annotation, and output.
• These hashes are provenance/cache-validity metadata. They describe the execution conditions, not the final result contents.
• If the spec has not changed, the UI reuses the latest compatible output folder and valid cached artifacts. In that case, no new execution folder is required.
• If the spec changes in a way that invalidates one or more stages, the UI creates a new timestamped execution folder.
• When a new execution folder is created, Autonima copies forward only cache artifacts that are still valid for the new config.
• Invalidated stages are rerun; valid upstream/downstream artifacts may be reused when safe.
• Each output folder stores the exact executed config snapshot and execution manifest for provenance.

For meta-analysis:

• You select a specific NIMADS source, which corresponds to a completed screening output folder containing outputs/nimads_studyset.json.
• The meta-analysis is run using that selected screening output.
• The resulting meta-analysis artifacts are associated with that selected NIMADS source, so different screening outputs can have separate meta-analysis results.