The whole engine on a small board
Raspberry Pi Zero 2 W. 6.8 MB engine, 36 ms answers and a valid flight mission, with nothing sent off the board.
Evidence
Datasheets: edge, integrity and learning.
Small hardware, grounding and learning, each with its datasheet and its data.
Against a local language model
Wikidata, 49 questions. 469 times less processor time for the same answers, with no fabrication from either.
Checking that a source exists
Wikidata, 78 titles. None of 20 invented titles given an author; 42 of 58 real works answered.
A real person at the camera
FaceForensics++ RAW. The altered face localised in 7 of 7 pairs, with none of 1,134 untouched regions changed.
Companion power on a drone
Simulation, three languages. The board costs 0.6% of the flight pack per flight; three independent implementations agree to the byte.
Mathematics learned from change
Held-out problem sets. No wrong answer across about 3,600 held-out problems; it declined where it could not compute exactly.
The showcases put several of these in front of a real Australian problem: showcases.senuaai.com. Security experiments are on the first datasheet page.