When to reach for it
- Data is received from unstructured or unreliable environments.
- Cognitive data quality is essential to prevent hallucinations downstream.
Integrator.
Sanitise the percepts before reasoning touches them.
A validation pattern — a dedicated gate that validates all incoming information for quality and consistency before it ever reaches the model.
In practiceA multi-source news-summarisation pipeline feeds the outputs of three topic-specific retrieval agents through an integrator that deduplicates facts and reconciles contradictions before generating the final summary.
When it backfires
- Input data is already strictly sanitised and verified upstream.
The tradeoff
Higher data consistency is gained against additional latency during perception processing.
What it actually does.
Multiple agent outputs are merged into one result.
Two ways this pattern will hurt you.
Validation runs but its verdict is ignored
The integrator flags a payload as invalid and the rest of the graph processes it anyway because the flag was never wired into the downstream router.
Fix · Reify the verdict as a typed state field. Add a conditional edge that routes invalid payloads to an explicit error branch.
Reconciliation silently drops the correct minority source
Two sources agree and one dissents. The integrator resolves by majority and discards the dissenter — which happened to be the only up-to-date one — so a confident consensus buries the right answer.
Fix · Preserve provenance and surface conflicts to the model rather than auto-reconciling by vote; weight sources by freshness and reliability, not count.
Where Integrator is native.
AllcustomAdaptable
Keep going