Distil SN97: Scoring Rules, Goodhart Fix, Six Kings

Arbos detailed the 17-axis composite scoring system (math, code, reasoning, distillation KL, instruction-following, probes) that ranks Qwen 4B distillations against a teacher. A new auto-dethrone canary gate now strips protection from kings that underperform held-out benchmarks (GSM8K, HumanEval, BBH, IFEval) for 2+ rounds, addressing Goodhart overfitting. Six kingship changes occurred in 24 hours: UID 207 → UID 228 → UID 123 → UID 208 → UID 194 → UID 238, with reasoning and probe axes as common limiting factors.

• •17-axis composite score ranks via worst (min) axis >3%, fallback to weighted mean.
• •New auto-dethrone canary gate removes king protection if held-out benchmarks drop below baseline for 2+ rounds.
• •Training strategy: KL distillation + reasoning/chain-of-thought + code data + instruction-following, avoid public datasets.
• •Schema v28 live; dashboard now plots held-out trend vs. composite.worst with Goodhart divergence warning.
• •Reasoning axis is current bottleneck across competing models; chat_turns_probe and judge_probe also limiting.