Researcher onboarding

AIPLA researcher guide — R1

For researchers

A researcher account layers cross-teacher, read-only views and a rubric experimentation workspace on top of the normal teacher surfaces. You observe across every teacher and class; you do not edit their work. Researcher access is a role an administrator grants to your account (see Getting access below).

What a researcher can do

Two capabilities beyond a normal teacher account:

1. Observe across the whole cohort — every teacher’s classes, activities, engagement and cost, read-only.
2. Experiment with rubrics — author and version the judge prompts (lenses) used to score student sessions, and run a judge over a captured session.

Everything a researcher sees across teachers is observation only — nothing in the cross-teacher views is editable.

Getting access

The researcher role is a claim an administrator sets on your account (via aiplatform users grant-researcher <uid>). Once granted, sign out and back in so your session picks up the new role. You will then see a Research item in the sidebar and extra controls described below. Without the role, these surfaces show a “Researcher access required” message.

Cross-teacher observation

• Research (activity scan) — the Research sidebar item opens a read-only scan of every teacher’s activities, in every state (Draft / Private / Shared), with the owner shown. Use it to see what is being built across the cohort.
• Research view (classes) — on Classes, a My classes / Research view toggle switches the list to every teacher’s classes; pick one to drill into its sessions.
• Insights across all teachers — on Insights, a My classes / All teachers scope toggle widens the engagement comparison to the whole cohort.
• Cost — a researcher-only Cost tab under Insights breaks cross-class spend down by cohort, model, voice (speech-to-text / text-to-speech) and class, over a period you choose.

The cross-teacher research surfaces are read-only — observation across every teacher and class.

Rubric experimentation

Under Settings, the Research · judge lenses panel is where you shape and test how student sessions are scored:

• Judge lenses — for each lens you can enable/disable it, choose the judge model, inspect the default prompt, and write a judge prompt override. Saving an override bumps its version, so you can iterate on the wording and keep the history. Reset to default reverts it.
• Experiment — score a captured session — paste a session id, pick a lens, and Run judge to see the score profile that lens produces on real data (or an “Abstained” result when the lens declines to score).

The rubric experimentation panel: versioned judge prompts and a judge you can run on a captured session.

This is the workflow for developing the capability-floor rubrics: change a lens prompt, run it against known sessions, compare the profiles, and promote a version when it behaves.

Next steps

• The teacher-facing surfaces work as documented in T1–T4; the researcher views sit alongside them.