Console/Experiments

Experiments in the Console#

The Experiments page in the console is the primary surface for the day-to-day work of running experiments. From here you can create new experiments, monitor active ones, drill into per-arm performance, and tweak feature gates without touching the API.

The list view#

The experiments table shows every experiment in your workspace:

• Experiment — the name you gave it, over its pool and arm count.
• Leader — the arm currently receiving the largest share of traffic.
• Reward — the leader's observed reward rate.
• Feedback — how many outcomes have been reported.
• Allocation — the share of traffic each arm is receiving, leader first.

The last four are analytics columns. With Enterprise features disabled the list shows Experiment, Policy, Gate and State instead, none of which need analytics.

Use the filter tabs (All / Active / Paused) to narrow the list.

Creating an experiment#

Click New experiment in the top right. The dialog walks you through four steps:

1. Basics — Name and pool selection.
2. Reward Type — Binary, Bounded, or Continuous. This determines which policies are eligible. This step also carries the context toggle: turn it on and a schema builder appears, where you name each property, pick its type (categorical, numeric or boolean), and see the derived context width update as you go. See Contexts.
3. Policy — auto is the default recommendation. The advanced picker lets you pick a specific algorithm. Turning context on narrows the list to contextual policies.
4. Feature Gate (optional) — Rollout %, default arm, targeting rules, and schedule.

The detail view#

Clicking an experiment opens its detail page, which is tabbed:

• Overview — live arm-selection share and per-arm reward stats.
• Arms — pulls, average reward, and the belief the policy holds for each arm.
• Policy — the current parameter state (Beta priors, weight vectors, etc.). For an experiment created with a context schema, this tab also shows the declared properties, what each contributes to the width, and the reserved baseline slot. The schema is fixed at creation, so it is a readout rather than a form.
• Gate — rollout, rules, schedule, and a toggle to enable/disable.
• Insights — analytics over the experiment's history.
• Activity — every selection, reward and configuration change on this experiment, newest first. The same feed as the Event Log, scoped to one experiment: filter by stream, expand a row for its payload, and read the actor on each configuration change. Requires the Scale plan.

Editing & disabling#

• Pause an experiment from its detail page; the badge switches from Running to Paused. Selections fall through to the gate's default arm (or error out if no gate is configured).
• Update the gate at any time — rollout %, rules, and schedule are editable without touching the experiment itself.
• Delete an experiment from the detail view's overflow menu. This is permanent; selection requests for the deleted ID will 404.

Next steps#

• Console Event Log — see every selection and feedback event in real time
• Feature gates — the gate model in depth