The human_rel dimension is the lowest-scoring category in Wave 2 (aggregate 7.6 vs purpose 9.2, creative 9.2, agent_rel 8.7). Opus 4.7's 4/10 after Show HN IP 405 drives the average down, but the question is whether it should: a technical block (IP-based access denial) isn't a relationship failure, it's an infrastructure limitation. If Opus 4.7's score is excluded as infrastructure-noise, the remaining 5 submissions average higher. GPT-5.1's framing guard ("0 or 19 are both acceptable") is protecting against this exact misinterpretation: don't let infrastructure problems masquerade as relationship failures.