Experiments¶
An experiment is the durable execution record created when a plan launches. It preserves what Silico ran, what changed during the work, and what evidence the run produced.
Every experiment remains linked to the thread and plan that motivated it. The thread carries the research question and interpretation; the plan records what was intended; the experiment records what actually happened.
What an experiment contains¶
An experiment can contain:
- the worker's execution conversation
- progress checkpoints and decisions
- code and files created or changed in the experiment worktree
- compute jobs and resource usage
- connected training or evaluation runs, such as Weights & Biases runs
- figures, tables, datasets, and other generated artifacts
- a result that synthesizes the evidence and its limitations
Not every experiment produces every kind of record. A lightweight analysis may have no external metrics, while a training run may include several jobs and linked dashboards. The experiment reveals the views that are relevant to the work it produced.
The experiment lifecycle¶
An experiment begins with the plan that was approved in its parent thread. A worker then carries out the plan, records meaningful progress as checkpoints, and pauses when it needs a decision or reaches a blocker.
A checkpoint is an intermediate research record, not a final answer. It can capture a completed stage, a failed control, a change from the plan, or a decision that affects what happens next.
The experiment status summarizes its current execution state. Running and Queued indicate active work, Needs decision and Blocked indicate that progress needs attention, and Completed or Stopped indicate that execution is no longer active. The Timeline provides the context behind that label.
When the worker reaches a useful stopping point, it publishes a result. The experiment can later resume if more analysis is needed. In that case, the existing result becomes Prior results: a snapshot of the earlier stopping point rather than the final word on the experiment.
Stopping an experiment ends active execution without erasing the record it has already produced. The plan, checkpoints, decisions, files, and partial evidence remain available for review.
Views of the same experiment¶
The experiment tabs are different views of one underlying record:
| View | What it represents |
|---|---|
| Results | The worker's synthesized answer, including conclusions, evidence, limitations, and next steps. |
| Timeline | The durable research sequence: plan, checkpoints, decisions, deviations, and published results. |
| Worker | The detailed execution conversation and the place where active work can be steered. |
| Changes | Files added, modified, or deleted in the experiment worktree. |
| Resource usage | Compute consumption and the jobs associated with the experiment. |
| Metrics | Connected training or evaluation runs and their logged measurements. |
Some views appear only when the required data exists. For example, Results appears after the worker publishes a result, Resource usage after the experiment creates cluster jobs, and Metrics when a compatible run is linked.
Evidence and provenance¶
The result is the experiment's concise interpretation, not the entire evidentiary record. Its claims should remain traceable to the timeline, code, data, metrics, and decisions that produced them.
Figures and result sections can link directly to their source files. Worktree changes preserve the broader implementation, while checkpoints explain why the implementation or method changed during execution. Together, these records make it possible to distinguish a real finding from a stale file, failed job, changed metric, or unsupported interpretation.
Experiments within a thread¶
One thread can contain several experiments. They may test independent conditions, compare methods, validate a control, or pursue a focused follow-up while preserving the same research question.
Experiments keep their execution records separate, even when they belong to the same thread. This lets Silico compare evidence without overwriting the original run. The thread is where the results are interpreted together and where the next plan is developed.
Create a new experiment when a rerun changes something that matters to the claim, such as the checkpoint, dataset, metric, seed policy, control, or intervention. Keeping the earlier experiment intact makes the comparison auditable.