It refused a higher point group when the added operator cost too much extra error
in the scaling model, relative to the parent's. It was guarded by "H could not be
computed" and followed by the H refusal, so it could only ever act where H - the
one statistic that discriminates here - is blind. Over 1213 candidate rows on 148
datasets it is reached seven times, every one of them with H not a number, fires
twice on a single dataset, and changes no result anywhere.
What it does do is make a space group depend on the compiler flags. On that one
dataset a plain Release build refuses a correct P2_12_12_1 and reports P2, while
the same source built for x86-64-v3 does not, because the quantity it compares is
not a measurement at all - it is the bisection's own ceiling, over a parent whose
error term is larger than the signal it is added to. A decision that moves with
the flags is worse than one that is consistently wrong, and no bound repairs this
one: genuine promotions now measure up to 5.8 where the bound was calibrated at
1.62 against a twin at 2.1, so the two populations have swapped sides.
The case it was meant to catch - a merohedral twin too sparse for H - is already
answered below it by the added-operator R gate, whose reference is the global best
rather than the parent, so it still speaks where H does not.
No space group and no cell changes on 148 datasets: 100 open, 34 in-house, 14
private.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_011GxZqDiFP3KqriBhNdcR56
leonarski_fandClaude Opus 5
dc0d7e65d8