Tuatara Decide vs Jev on BANKING77: a tie at a thousandth of the cost, then a win

We scored a Tuatara Vector Model blend against Jev’s published answers on all 3,080 BANKING77 test messages: 91.79% versus 92.40%, a statistical tie, at about 14 cents of CPU per million messages against about $142 in API charges. Updated September 28: an embedding head trained on the same messages now scores 94.09%, ahead of Jev (p = 0.00014), at about $5.70 per million. Jev still leads with no training examples.