Distil SN97: System Breakdown Amid King Selection Chaos

Over 24 hours of Discord messages from the Distil subnet team reveal critical infrastructure failures undermining fair competition. King selection logic is backward or broken (lower-scoring models dethroning higher-scoring ones), state persistence is non-functional, and single-eval mode creates unfair prompt-variance comparisons. The max_new_tokens bug truncates teacher responses to 1024 tokens (not 8192), corrupting all evaluation data. Multiple false DQs (including A_Tensor's model wrongly flagged as a copy) suggest validator bugs. Miners can't see composite scores—only KL announcements—creating opacity around selection criteria. Community has extensively documented these issues; const (operator) has acknowledged none of them publicly.

• •King selection appears inverted: UID 93 (lower score) dethroned UID 162 (higher); UID 149 beat UID 148 despite worse KL on different prompt sets—single-eval mode unfairness.
• •max_new_tokens capped at 1024 (not 8192)—teacher responses truncated, all composite scores evaluated on corrupted data, violates evaluation freeze.
• •Copy detection logic backwards: A_Tensor's model (committed block 8051983) DQ'd as copy of ivangrapher's (block 8052524), despite earlier commitment.
• •h2h_latest.json state persistence broken; king identity not saved between rounds; validator precheck comparing against ghost revisions (UID 169, 174).
• •Composite scores hidden from miners; only KL shown; no transparency on what miners must beat; miners can't calibrate locally.