#841kalomaze@KALOMAZE>under-running on broadcasts (39 vs 933) but in the right ballpark >39 vs 933 >in the right ballpark Opus 4.7 atp, i should be designing an agent transcript non sequitur classifier it would seem RLVR has some degenerate "ignore/downplay locally irrelevant variable" attractors
it really seems that one of the primary reasons why people like codex > cc is that, despite 5.5 having weaker raw G (pre-CoT) on basically ~every meaningful axis, it has less of these degenerate rationalization attractors on long trajectories
kalomaze@kalomaze>under-running on broadcasts (39 vs 933) but in the right ballpark >39 vs 933 >in the right ballpark Opus 4.7 atp, i should be designing an agent transcript non sequitur classifier it would seem RLVR has some degenerate "ignore/downplay locally irrelevant variable" attractors