How Silico organizes research¶
Silico is organized around the shape of a research project, not the shape of its backend. You only need four ideas:
Thread → plan → experiment → report
The thread is the continuing investigation. Plans make proposed work reviewable before it starts. Experiments preserve what actually happened. Reports turn the evidence into something another person can understand and use.
Follow one question through the system¶
Suppose you want to know why a model hallucinates on a particular class of prompts.
You start a thread with the question, the model, and a representative dataset. Silico proposes a plan for a baseline, an internal measurement, and a held-out evaluation. Launching that plan creates an experiment. Its Timeline records the work and its Report explains what the evidence supports.
The first result may narrow the question rather than settle it. You can ask for a controlled follow-up in the same thread. That creates another plan and another experiment, while the thread preserves why the follow-up exists and how it changes the interpretation.
This separation matters months later. You can read the thread to understand the research direction, open an experiment to audit execution, or share a report with someone who only needs the result.
Know when to make something new¶
| Keep using… | When… |
|---|---|
| the same thread | the next question depends on the reasoning or evidence already there |
| a new plan | the thread is ready for another concrete piece of work |
| a new experiment | a rerun or changed condition deserves its own execution record |
| a new report | the output answers a distinct question or needs to serve a different audience |
Start a new thread when the question can stand on its own. Put related independent threads in a folder rather than forcing them into one research history.
Context belongs to the moment that needs it¶
Context is not another container to manage. Files, repositories, papers, data, models, and earlier work stay attached to the thread message that needs them.
That placement is part of the record. A dataset attached for a follow-up does not silently become the input to every earlier experiment, and an old report does not become evidence in an unrelated thread unless you attach it.
Personal and organization settings hold the things that really are reusable: credentials, integrations, runtime variables, compute, and standing instructions.
Read the layer that answers your question¶
- Open Summary to learn where a thread stands now.
- Open Timeline to reconstruct how it got there.
- Open an experiment to inspect its Plan, Report, Code, Compute, or Metrics.
- Open Shared Reports when you remember the result but not its source thread.
When you are ready to put this model into practice, continue with Working with a thread.