Validated skills are necessary but not sufficient for trustworthy agentic genomics¶
Manuel Corpas, Crick AI Club, Francis Crick Institute. Recorded on Zoom, 14 September 2026.
Science after pipelines. This talk asks what changes when scientists direct agents through objectives instead of manually connecting computational tools. It introduces ClawBio's inspectable scientific skills and the separation between reasoning, specification and execution, then uses two pharmacogenomics figures to examine what validated rules improve and what remains outside their guarantee.
Download the PowerPoint with speaker notes
Watch the full talk on YouTube
Three ideas to remember¶
- Scientists should be able to direct scientific inquiry through objectives while retaining judgement over methods and evidence.
- Explicit, versioned skills make scientific methods inspectable. Deterministic execution adds reproducibility and an auditable decision boundary.
- Validating one component does not validate the whole system. Inputs, tool selection and the connections between skills need their own evidence.
Evidence in the talk¶
- A controlled pharmacogenomics comparison shows the effect of supplying validated rules on recommendation accuracy.
- A separate input-interpretation experiment shows why coverage and accuracy must be examined together, even when allele definitions are supplied.
- Figures 2 and 4 from manuscript v90 connect the architecture to these measured results. The high-severity error counts depend on the scoring rule and are not measurements of patient harm.
The study evaluates components, not an end-to-end autonomous clinical system. The recording is 34 minutes 13 seconds and includes the talk and closing remarks; the announced Q&A and possible demo are not included. The downloadable deck contains 22 slides, including two appendix slides. Repository figures in the slides are dated snapshots.
Benchmark data and provenance | Event page
Read the documentation | Follow the events calendar | Join the ClawBio Slack channel