Run your first experiment¶
Your first useful Silico result begins with a question, not a configuration screen. This walkthrough follows one investigation from an initial idea to evidence you can inspect, build on, and share.
We will use hallucination prediction as the example, but the same rhythm applies to interpretability, model evaluation, steering, data analysis, and tool building:
Ask → plan → run → inspect → follow up
You can use the defaults for a first run. The important part is learning where to express intent and where to verify what happened.
1. Start a thread¶
Choose New thread and describe what you want to learn. A clear question is enough; Silico can help turn it into a method.
I want to understand whether a small set of internal features can predict when this model hallucinates.
Then add the details that would change the answer: which model, what data, how success should be measured, and any practical limit on compute or time. Attach a model, dataset, repository, report, or paper when it is part of the question rather than merely background reading.
You might continue:
Use the attached evaluation set, keep its held-out split untouched, and produce a short report that shows whether the signal transfers across domains.
Enable Auto-accept only when you are comfortable letting routine work proceed after a review window.
2. Review the plan¶
The plan is your chance to slow down before work begins. Read it for the mistakes that could survive all the way into a convincing report.
Confirm the model and checkpoint, dataset and split, metric, controls, expected report, compute footprint, and stopping conditions. Open an input chip whenever identity or provenance is unclear.
Comment on a passage for a local correction. Reply in the thread when the plan reveals that the overall question or direction should change.
3. Launch and follow Timeline¶
Choose Launch experiment. The plan becomes a separate experiment, and Timeline begins recording meaningful progress, decisions, deviations, and reports.
You can leave the experiment running and return later. You do not need to watch it continuously. If Silico needs a decision, it appears in Timeline with the surrounding context.
For example, if a control fails, answer with the smallest constraint that preserves the purpose of the run:
Stop before the full sweep. Debug the control on 20 examples and keep the current outputs so we can compare them after the fix.
Change research direction through the thread; use Worker only for an immediate execution correction.
4. Inspect the report¶
Read the report once as a reader: what does it claim, what evidence matters most, and where does it stop being reliable? Then read it as a reviewer. Follow source links into Timeline, Plan, Code, Compute, or Metrics where the decision depends on the exact implementation or data.
For this example, you would want to know whether the probe was evaluated on held-out data, whether transfer was measured on the same checkpoint, and whether a simple behavioral baseline explains the result just as well.
When the experiment completes, start with Reports for the conclusion. Use Timeline to understand how the work developed and Code to inspect the files behind it.
Copy the report link when someone needs the output. Copy the thread link when they need the full research record.
5. Continue the loop¶
Return to the thread's Summary to see where the research now stands. A first experiment often replaces a broad question with a sharper one. That is progress.
Ask for the smallest follow-up that would change your interpretation. Keep it in the same thread when it depends on this evidence, but launch it as a separate experiment when its conditions or claim differ.
A good loop ends with an answer, supporting evidence, the main uncertainty, and a next decision. The report makes the result portable; the thread makes the reasoning durable.