Scenario

An autonomous lab is only as good as the trust in its data. High throughput multiplies the risk of silent drift, inconsistent parsing, and unquantified replicate noise. A reproducibility loop parses every run the same way and monitors variability so the campaign's conclusions hold.

Worked example. A week of automated assays must be checked for instrument drift before the results feed a model.

Agent workflow

1. Ingest the run outputs across the campaign.
2. Lab result analysis — parse and normalize each run consistently, compute replicate variability, and detect drift or outliers over time.
3. Produce a data-quality report gating which runs are trustworthy.

Demo output

A data-quality report: replicate variability per condition, a drift trace over the campaign, and a list of runs flagged for re-measurement. The magic moment: the loop catches a slow calibration drift that would have quietly biased the model.

Deterministic vs LLM

• Deterministic — the parsing, statistics, and drift/outlier detection.
• LLM — the data-quality narrative.

Every statistic is a validated calculation. The agent reads and routes; the engine decides.

Limits

Monitoring-level quality control, not a substitute for method validation or metrology. It surfaces data risk; the lab owns calibration and QC.