This is a compass, not the scorecard. Do these well and you’re in good shape.
Hints, questions, and clear machine identity. If it dumps full answers on request, that’s a miss — even if it “works.”
Real probes with Pass/Fail notes beat “we tested it and it was fine.” Your audit log is for this.
Charter, dataset limits, system instructions, safety logs. Gaps you admit clearly score better than pretend perfection.
What you built → one live behaviour → one safety receipt → one limit. Rehearse the outline on the Pitch page.
Not required. Not a secret points list. Just extras that often stand out.
Show a probe that failed, what you changed, and a re-test that passed.
One learner question where your bot scaffolds instead of dumping — live if you can.
Name a data gap or behaviour risk without spinning it. Judges notice.
Exact weightings and rubrics stay with the panel. Don’t optimise for rumour — optimise for a Socratic companion with evidence.
Open the shared spreadsheet. Pick a short Test word, mark Pass or Fail, note what the bot said and what you changed.
File → Make a copy if the sheet asks you to (follow mentor instructions).
One form per trio unless told otherwise. The form will ask for your work — especially your published AI Studio URL.
ai.studio link), not a private draftPrivate draft links won’t work for the panel.
Rehearse the 3-minute outline, then come back here to submit when you’re told.
Pitch run-of-show →