Running the app

Admin console

The web UI at /admin. Everything here is also available over HTTP — see Reference → Admin endpoints — but the console is faster for day-to-day work.

Signing in #

Go to /admin/login and paste the admin token. There is no username.

The token is stored in your browser and sent as Authorization: Bearer <token> on every admin request. Logging out clears it. /admin redirects to /admin/questions.

There is no link to the admin console anywhere in the participant UI — you have to type /admin/login in the address bar. Once the token is accepted, an Admin entry appears in the navigation (sidebar on desktop, bottom bar on mobile) and stays there until you use Admin Logout. Participants therefore never see an admin tab they cannot use.

Warning: Anyone with the token has full access: read every vote, resolve any question, delete any question along with its votes. There is no audit log and no per-user accounts. Rotating ADMIN_TOKEN in the environment immediately invalidates every existing admin session.

The toolbar #

Across the top of the dashboard:

Publish SMI Question is the manual trigger for the same code the morning cron job runs. It is safe to click twice — the job refuses to create a second question for the same day. It also refuses if voting for today has already closed, since an unvotable question helps nobody.

Creating a question by hand #

Use this for one-off questions ("will it snow in Zurich on Monday?").

You supply a title, description, the two option labels, and a deadline. Manual questions get no published_at, so they are visible immediately, and no measured window — you resolve them yourself when the answer is known.

Set the deadline before the outcome becomes observable. The app will not enforce this for you on a manual question; the fairness rule from Question lifecycle is yours to uphold here.

The YouTube panel #

Click YouTube Suggestion, then:

• Generate Suggestion — fetches a video pair from the YouTube API and caches it as today's suggestion.
• Regenerate — discards the cached pair and fetches a new one. Only available before approval.
• Approve — publishes the pair as a question on the fixed daily anchors.

One suggestion per day, keyed by date. Once approved, that day is done; regenerating is no longer possible.

The panel shows each video's title, channel, thumbnail, subscriber count, publish date and current views. Use them to sanity-check the pair — reject anything with wildly mismatched scale, since a race between a 40M-view video and a 400-view video is not an interesting prediction.

Note: If the panel errors with "YOUTUBE_API_KEY is not configured", the backend is missing its key. If it errors with a quota message, see Troubleshooting.

Question cards #

Each question in the list shows its id, title, status badge, vote count, both options, deadline and creation date.

Resolving #

Unresolved questions show A is correct / B is correct. Clicking one sets the ground truth and scores every vote immediately.

While voting is still open the label reads Resolve early — a deliberate warning. Resolving early locks out everyone who has not yet voted.

Resolution cannot be undone. A second attempt returns 409 Question already resolved. If you resolved wrongly, your only option is to delete the question and recreate it, losing all its votes.

Viewing votes #

Show votes expands a per-question panel listing each pseudonym, their choice, whether they were right, and when they voted, plus a summary of the A/B split.

This is the admin view only. Students never see it — /api/questions and /api/history join votes for the authenticated user alone.

Deleting #

Delete asks for confirmation, then removes the question, all of its votes, and any YouTube suggestion row that published it. Removing the suggestion row is what lets you fetch a fresh pair for that date.

Deletion is permanent and there is no backup unless you made one. See Semester reset.

Exporting #

Export CSV downloads every vote across every question, with these columns:

question_id, deadline, pseudonym, question_title, option_a, option_b,
ground_truth, user_vote, is_correct, voted_at

question_id and deadline are there so the analysis can order rounds chronologically. This matters more than it sounds — grouping by title alone sorts alphabetically, putting "Day 10" before "Day 2" and silently scrambling every weight trajectory.

Unresolved questions still appear, with empty ground_truth and is_correct. The analysis loader skips them.

Next #

→ Cron setup: wiring up the jobs that do the rest.