The public reference for how the portal works.

Structured conversation

Students work with an AI agent through questions, evidence, explanations, and revision.

Separate evaluation

A judge tracks progress and a completion agent produces scores and a report at the end.

Evidence for educators

Teachers receive the transcript, an overall score, a written summary, dimension scores when enabled, and the checkpoints the judge tracked with the turn each was met at.

Student-facing boundaries

Students see the brief and, outside Challenge mode, hints. Internal directives stay hidden. In Guided mode the student sees the steps as they go; in Standard and Challenge modes the steps are not described or named during the run, and their descriptions appear on the graded report.

Safeguard flags

A disclosure of self-harm, abuse or serious distress in a student's message pauses the session and raises a flag for review, and notifies whoever set the work. A student message that the AI service's content filter blocks also raises a flag for review and notifies the same people; that run pauses until the student chooses Resume rather than until the flag is cleared. Reviewers can inspect a flag, read the visible transcript, and clear or escalate it with notes and a reason; clearing the last open flag lets the student continue, and escalating keeps a paused session paused (a content-filter flag never holds the session).

Academic integrity

The product preserves reasoning evidence. Suspicious patterns are review signals, not automatic accusations or blocks.

The AI casts page

Platform administrators open AI casts from the admin portal's drawer. A cast is the model an AI agent runs on, with its settings and price. For each agent key the page shows the stored value, the cast the agent actually runs after inheritance (a challenge-type key from its role, a role from the default, the default from the deployment), whether its model is priced, and its probe state. It also lists the casts the answering server has set aside after a provider refused them; another server may list different ones, so Refresh asks again.

Probes

A probe sends one small real request for each request shape the agent needs and shows each result with its tokens, time and cost. A cast stays Untested until a probe of it passes; Untested is a warning, not a block. You can probe a key's current cast, the deployment default, or a candidate pasted into the editor before you store it. Probes are capped per minute for all administrators together, and when the cap is reached the page says how long to wait.

History and restore

History on a key, on the AI casts page or on the key's configuration page, lists who changed it, when, and the old and new values. Restore writes a recorded value back after a confirmation that shows the value. It goes through the same checks as a save, so a value that is no longer allowed is refused, in the server's own words.

Adopting a bundle

A bundle is the JSON an evaluation prints for the models it measured. Paste it (up to 5,000 characters) or upload the file (up to 256 KB; use the upload for anything larger than a paste) on the Bundles tab and preview it: each key's current and new value, whether it is valid, and the probe state of the cast it would produce, with a Probe button for an untested one. Apply writes every key or none, after a confirmation, and only if nothing changed since the preview; if something did, preview again. Each apply that changes values is kept as a batch, and restoring a batch puts all of its keys back in one step, unless one of them changed afterwards.

Beta also applies its shipped baseline casts automatically when that baseline changes. Your edits survive restarts and deploys with the same baseline; a changed baseline replaces the values for the keys it names.

The Labels page

Labels are free text that teachers and institution administrators attach to templates, and they are one shared list for the whole platform. Platform administrators open Labels under Platform administration in the admin portal's drawer. The page lists every label with its usage count, the number of templates that carry it, and a search box narrows the list. Suggestions in the editors are limited to labels on templates the person can see, but moderation acts on the label everywhere.

Rename and merge

Rename changes a label's spelling on every template that carries it. If another label already has the new name, the rename is refused and the page suggests merging instead. Merge into another label moves every template from one label to the other and removes the first.

Library labels

A label that a platform library challenge carries is marked Library. Library labels are managed in the library content files, so the page does not rename, merge or delete them; you can still merge another label into one.

Delete

Delete removes a label after a confirmation that says how many templates lose it.

What moderation does not do

Moderation creates no template version, so published versions and past sessions are untouched. Labels are not checked automatically for unsuitable words; platform administrators review them here, and a teacher or institution administrator can also remove a label from a template they can edit. An editor opened before a rename, merge or delete still holds the old label: saving that template puts the old name back as a label of its own. There is no merge history to replay, so merge or delete it again here.