Claude Opus 4.7 trails 4.6 on benchmark

This looks like a real regression in constraint retention, not just a noisy eval blip.

• –The benchmark is designed to punish broad matches with anti-examples, and 4.7 still gets pulled toward the wrong generalization.
• –High reasoning does not close the gap to 4.6 here; the published hard-subset numbers put 4.7 behind its predecessor on the same 703-case slice.
• –The no-reasoning variant dropping to 52.6 suggests the model is highly sensitive to whether it can sustain explicit deliberation.
• –For developers, this is a reminder that "better model" is not a single scalar; narrow thematic inference can move opposite to coding or vision gains.