SN97 Scoring Overhaul, Validator Downtime Ongoing

Distil rolled out v32 composite scoring across 23 axes, replacing v31's tighter token limits and changing the weighting scheme: on-policy RKL (0.39), long-form coherence (0.25), and judge probes (0.20) now dominate. Three models dethroned the king in 24 hours (UID 238, 214, 177). Validator pod remains stuck on a tokenizer cache issue with Kimi-K2.6; miners report evaluation backlog and are pressing for urgent fixes.

• •V32 scoring: 23 axes, worst-3-mean formula 0.75×worst_3 + 0.25×weighted; dethrone margin 3%
• •New axis: long_gen_coherence (0.25 weight) measures structured long-form text without teacher model
• •Validator offline due to corrupted HuggingFace cache on Kimi-K2.6 tokenizer; ~290 models DQ'd
• •Three crown changes in 24 hours; current king UID 177 earns ~$5,700/day from pool of ~$89.50/day
• •Token limits flagged as too tight on v31 knowledge, long-context, and calibration benches