01 · Output quality
8/10
questions won
Same model, two data layers. A blind judge scored Amdahl higher on 8 of 10 GTM questions.
02 · Consistency
73%
of numbers repeated, vs 28%
Same question, five runs. 73% of the numbers came back the same, against 28% without Amdahl.
03 · Token efficiency
400x
fewer tokens, worked example
200 calls is about 800K tokens of transcript. The cluster an agent needs is about 2K.