Resources and artifacts¶
Resources and artifacts connect Silico's conversations to the concrete material used and produced by research.
Resources¶
A resource is context Silico may use while planning or running work. Resources can include:
- GitHub repositories and code folders
- models and checkpoints
- datasets, evaluation sets, and example cohorts
- uploaded files
- cloud storage locations
- outputs from earlier experiments
Make a resource reusable when several threads should be able to find it. Attach it directly to a thread when it matters only to the current question.
A resource link tells Silico where relevant material exists. If the material requires special handling, explain that in the thread or in stable organization instructions.
Attaching a resource adds context; it does not automatically copy every byte into the conversation. Access still depends on the relevant integration, environment variable, filesystem, and compute fabric.
Artifacts¶
An artifact is a durable output of an experiment: a result file, table, diagram, checkpoint, trained probe, evaluation output, report, or other generated evidence.
Artifacts may live with the experiment or in connected cloud storage, depending on the compute used. Silico keeps links between the result and the artifacts that support it.
Prefer durable artifact locations for outputs that must survive a worker restart or be reused by another experiment. A temporary file in a worktree is useful during execution but is not a good handoff until the result or Timeline points to it clearly.
Inputs and outputs¶
The same file can play both roles over time. An artifact from one experiment can become a resource for a follow-up thread. Preserve that provenance: attach the prior result or artifact instead of copying a value without its context.
When reusing an artifact, record the producing experiment, version or checkpoint, and any transformation applied before the new run.
What belongs in the result¶
Results should summarize the conclusion and point to the supporting artifacts. Large data, model checkpoints, and detailed logs should remain artifacts rather than being copied into the narrative.
This separation keeps the answer readable while preserving enough evidence to inspect, reproduce, or extend the work.