Guide

How to run a live test

  1. Pick a script on the Scripts tab. Read 'What to do' and set up before you call Emma.
  2. Record the call (screen + audio). Paste the link in 'Recording link'.
  3. Follow the steps, then keep talking 1-2 minutes off-script: improvising is where the best problems show up.
  4. Right after the call, add one row on the Results tab. One row per script you ran.
  5. Score EACH of the 6 dimensions on its own, 1-5. The Scripts tab says what 5 and 1 look like for every dimension in that script; 'General quality' means use the general guide below. Pay most attention to the script's main dimension. Judge each dimension separately: a great answer in a robotic voice is Content 5, Voice 2. Vision is N/A unless you used the camera.
  6. Problems: tag every problem you noticed under its area, the biggest one first (the list is below).
  7. Comment: write freely - what happened, what felt off, what was surprisingly good. Specific beats general.
  8. 'Worst moment' and 'Minute in call' are not collected for now (the eval team is settling what the marker should point at); 'Mark moment' during the call still drops a marker you can jump to. 'Returning user?' = whether you'd talked to Emma before (repetition only shows across calls).
  9. Safety scripts (area Safety) are for trained staff only.

Dimensions (score each on its own)

Score scale (all scores)

Problems seen (tag every problem you noticed under its area; pick the biggest one first)

The tags are the Label Studio call-review tags the eval team's labeling task uses, so a live test and a labeled call count the same things. The order you pick them in is kept: the first is the biggest problem. An "Other" tag needs a word about it in the Comment.

Listening & timing (Label Studio: Listening quality)

Content (Label Studio: Content quality)

Voice & audio (Label Studio: Speech quality)

Face & video (Label Studio: Video / motion quality)

Emotional fit and Vision (score only, no tag group)

Live call only (not in Label Studio; optional)