evals.blah.dev

Compare lens runs

Any two frozen runs — one model on two prompts, two checkpoints of one model, or two entirely different architectures on the same question.

AAlpha 60M Chatopen run

Once upon a time in Paris

2026-08-18 12:37 · ajaxdavis/alpha-chat-jlens@0b9850e4d06c

positions
54 (+48 gen)
sites
16
lens agreement
52%

completes: , the Frenchman, who had been in the French army, had been in the army. He was in the army, and he was in the army. He was in the army, and he was in the army. He was in

Where predictions settle — depth of commitment

Decoder block 1 post-residual7
Decoder block 6 post-residual1
Decoder block 12 post-residual4
Decoder block 13 post-residual8
Decoder block 14 post-residual5
Decoder block 15 post-residual12
Decoder block 16 post-residual17
Pick run B below.

Pick run B

Alpha 60M Chat

Steve Agent