• How do you decide the level of metadata required for Vivarium models?

• Does Vivarium follow MIRIAM?

We require metadata that supports execution, composition, interpretation, and reproducibility: typed inputs and outputs, units, configuration, update behavior, dependencies, version, provenance. Additional metadata should capture the model’s assumptions and intended reuse​. Not following MIRIAM.

• What other groups are developing vivarium models, or converting their models to vivarium? 

• What kinds of meetings do you/your lab attend? 

• For modeling

• For standards and FAIR?

• Could vivarium be used for the individual and population scales, not just cells/tissues? What would that take?

• Are their lessons that we can take from the bigraph/compartmental hierarchy that might extend to multiscale models that use many distinct agent types?

The hierarchy separates what an entity is, where it is, and how it interacts. Distinct agent types can share common interfaces while retaining specialized internal models. This enables heterogeneous agents to coexist, move between compartments, exchange information, and be added or removed dynamically. The Bigraph framework is a generalized approach to modeling “the space and motion of communicating agents"

• How do we think about communicating uncertainty when uncertainty at one scale can magnify at another scale?

Uncertainty should be represented at each model interface, propagated through ensembles or probabilistic distributions, and reported locally between interfaces and for final outcomes. 

• The platform we are building may need to inform responses to future outbreaks. What validation, versioning, provenance, or compatibility checking may we want to consider as we develop the platform if there is the potential for bad models to be combined with good models? 

Every component should have a versioned contract, provenance record, validation evidence, applicable domain, units, and assumptions. Composition includes automated checks for interface compatibility, units, timescales, semantic meaning, parameter ranges, and validation status

• As a disease platform, we may need to host/run large models. How might we want to think about that now in the early stages of development? 

Design early for containerized components, data references, checkpointing, parallel execution, cloud/HPC scheduling, resource declarations, and reproducible environments. Models should run locally when small but move to distributed infrastructure without changing their scientific interfaces. Vivarium helps with all of this.

• Where and how in vivarium 2 are:

• Agents used

• LLMs used

Agents are used in surrounding tools to help wrap simulators, generate process contracts, assemble composites, search documentation, explain models, and operate the Vivarium Workbench. There are no LLMs in the core engine. Viva-superpowers is the ai-specific repo that knows how to interact with vivarium: https://github.com/vivarium-collective/viva-superpowers

• What are the minimum requirements to run workbench with a setup similar to what was demoed? Do you have links to any documentation?

Here is some info about the software: https://vivariumlab.com/software/. 

I demoed the E. Coli whole-cell model in vivarium 2: https://github.com/vivarium-collective/v2ecoli. If you go through that readme there is a lot of information about how to run it, including a markdown file you can give to claude so it will get everything running for you: https://github.com/vivarium-collective/v2ecoli/blob/main/docs/first-run-agent-guide.md. Creating and running your own repo takes a little bit more, because you have to identify the components and templates that you need.