System note · biomedical

Circuit Compass

Decision-support loop

Do not collapse the paradox.
Make it testable.

The system accelerates hypothesis generation. Authority remains with the experiment that can disagree.

01

Begin with the contradiction

The starting point was not a dataset but a mismatch in a collaborator’s experiment: behavior and molecular readouts appeared to tell different stories. Instead of asking one model for an explanation, the system formalizes competing mechanisms and asks which public evidence would move each one.

That changes the output from a plausible narrative into a decision surface: mechanisms, expected directions, relevant cell types, supporting literature, and the next experiment that could discriminate between them.

02

Build an inspectable handoff

The pipeline ingests public single-nucleus datasets, performs mouse-level pseudobulk analysis when biological replication permits it, organizes signals by mechanism, and generates figures and a research handoff. Synthetic data remains available for installation checks; real-data outputs are kept distinct.

The useful artifact is not an answer key. It is a PI-meeting object that shows where the signal came from, which interpretation was applied, and which physical measurement should come next.

03

Attack the story before the meeting

A dedicated critique names the hardest weaknesses: unresolved sex-versus-strain confounding, single-cohort findings, partial circularity in cell annotation, assay compression, mRNA-to-protein distance, collapsed spatial structure, and alternative explanations for the same signature.

Those caveats are not a footer. They determine the next work: find matched cohorts, replicate in another dataset, change the statistical model, split anatomy, add another modality, and let wet-lab measurement arbitrate the mechanism.

04

The boundary is the point

Circuit Compass does not replace an experiment or establish a biological mechanism. It makes computational synthesis cheap enough to explore several explanations while preserving the reasons not to believe any one of them too quickly.

This is the same control problem as agent evaluation at another altitude: propose widely, keep provenance, expose the failure modes, and hand authority to the verifier that can touch reality.