Review and launch a plan

A plan is the pause between a promising idea and paid, state-changing work. It turns the conversation into a proposal you can inspect: what Silico will do, what it will use, what evidence it should produce, and when it should stop.

The point of review is not to rewrite the plan in your own words. It is to catch assumptions that would make a polished result answer the wrong question.

Review what matters

Open the plan beside the thread and focus on choices that would be expensive, scientifically important, or difficult to undo:

• Goal: Does this answer the question?
• Inputs: Are the model, checkpoint, data, split, repository, and paths correct?
• Method: Are the comparisons, controls, and measurements sufficient?
• Output: Will it produce the report, code, data, or checkpoint you need?
• Resources: Are accelerator, parallelism, runtime, storage, and cost appropriate?
• Stopping conditions: When should it stop, scale, or ask you?

Open input chips to verify version, scope, and origin. “The validation set” is not specific enough when several variants exist; a named dataset, revision, split, and path usually is.

A small exploratory run can stay lightweight. Ask for more detail when the work is expensive, sensitive, or likely to support an important claim. Rigor should scale with the consequence of being wrong.

Read the plan as a future reviewer

Imagine returning after the experiment has finished. Would the plan let you tell whether the result tested the intended model, used a clean evaluation split, included a meaningful control, and stopped for the right reason?

If not, add the missing constraint now. A good plan does not predict every implementation detail, but it makes the evidence auditable.

Give feedback in context

Respond in the thread when you want to change the overall direction. For a local correction, highlight the relevant passage and leave feedback there so the surrounding intent stays clear.

Prefer feedback that names both the correction and why it matters:

Use the held-out patient split for the reported metric. The current split was used during prompt development and would overstate generalization.

Launch the experiment

Choose Launch experiment when the plan reflects the work you want done. Silico creates a separate experiment in the thread Timeline so the approved intent, execution, and evidence stay together.

With Auto-accept enabled, the plan card shows a three-minute countdown. You can launch sooner, pause the countdown to keep reviewing, or let Silico launch when the window ends.

If a credential, dataset, model, or compute option is unavailable, Silico asks instead of quietly substituting another input. Resolve the missing dependency or update the plan explicitly; either choice then remains visible in the record.

The experiment opens on Timeline. The approved plan remains in Plan; Reports appears when one exists.

Next step

Continue with Monitor and steer an experiment.