Admin Console Overview

Researchers run studies through the admin console at /admin. This page walks the navigation, the experiment lifecycle, and what each tab does. Most tabs have a dedicated deep-dive page. The cross-links at the bottom of every section take you there.

Navigation

The admin console exposes three top-level routes.

Route	Purpose
/admin (Dashboard)	Quick view of every experiment plus aggregate counts (total experiments, active count, total users, total videos)
/admin/experiments	Full experiment list. Create, edit, or delete an experiment from here
/admin/ui-custom	UI template library (visual or code track) shared across experiments

Open an experiment from either Dashboard or Experiments and you land on /admin/experiments/<id>, which is the per-experiment workspace. The remaining tabs all live there.

Experiment lifecycle

An experiment moves through three statuses. The admin flips them manually with the dropdown at the top of the experiment page.

Status	What it means	What is allowed
draft	Setup stage. Edit groups, users, videos, UI configuration freely	Everything
active	Participants log in and stream events into the database	Algorithm and group config can still be edited (researchers iterate). UI config and device assignment are frozen so layout-driven confounds do not appear mid-study
completed	Study has ended. Data is read-only	Group, user, and video edits are blocked. Stats and exports still work

Post-study surveys (see Designing Surveys) trigger only when status='completed' AND the survey is is_active=true.

Per-experiment tabs

Once you open an experiment, seven tabs run across the top.

Overview

Three counters (user groups, total users, total videos). It is the landing tab and is intentionally light. The detail lives in the other tabs.

Groups

The list of user groups in this experiment. Each row carries a device class, an algorithm config (one key per surface, feed and watch), a UI config, and the user count.

Adding a group is one form. Pick a name, a device class, an algorithm pair, and a UI config. The platform validates that the UI keys you pick support the group's device, so a tablet group cannot silently land on a desktop-only built-in.

The Edit button on each row opens the algorithm config modal. UI config and device assignment freeze the moment the experiment goes active, so the modal switches to algorithm-only mode after that point.

Users

Per-experiment participant list. Each row carries the participant's login id, group name, last-login timestamp, and a Status button that opens a modal with the participant's full event trajectory grouped by day.

Bulk Create generates participants under a chosen group with random passwords. The credentials are shown once. Download the CSV before closing the modal.

Download CSV exports the participant list (without passwords) for record-keeping.

Videos

The video pool. Two import paths.

1. Manual CSV upload. The CSV header is documented in Adding a Recommender. Rows in the CSV become videos in the experiment.
2. Auto-import from a dataset folder. Drop a dataset folder under data/<name>/ on the host. The Auto-import button on this tab reads the videos CSV plus the comments CSV plus the videos and thumbnails subfolders.

Edit mode flips the table into bulk-select. Pagination handles larger video pools (100 per page).

Surveys

CRUD for the experiment's surveys. Three sections (Pre-study, Inter-session, Post-study) match the three timing kinds. The full data model and the priority dispatcher are documented in Survey System. The admin-side authoring flow is in Designing Surveys.

Stats

Aggregate metrics for the experiment. CTR, average watch time, watch ratio (median), session length (median videos), session duration (median seconds), and per-group comparison. The metric definitions plus the low-confidence threshold are in Stats Tab.

The Export Events CSV button on this tab streams every event in the experiment as a UTF-8 CSV with BOM (Excel-friendly). One row per event, with the policy and feed position attached. This is the artifact used for downstream analysis.

RecBole

Visible only when at least one group's algorithm config sets feed or watch to recbole. Shows training status, coverage, fallback stage usage, and training history. The full reference is in RecBole Tab.

What does not have a tab

A few admin actions are reachable but not tab-shaped.

• Recommender registry sits at /admin/recommenders (not yet tabbed). Use it to register an external HTTP recommender. See Adding an External Service.
• UI templates live at /admin/ui-custom and are shared across experiments. Each experiment references templates by id from its groups' UI config.

Where to go next

• Stats Tab. Read the metric definitions used in the Stats tab.
• RecBole Tab. Read about training cadence, coverage, and the fallback chain.
• Quick Start. Bring the platform up locally and click through the lifecycle once.