SN3 Debugging Eval Discrepancies, Infrastructure Updates

τeuτonic team spent the day troubleshooting a persistent gap between local and remote model evaluation scores, with dashboard-to-validator mismatches of 0.1–0.2 loss. Infrastructure changes included spinning up a more consistent machine and testing Lium/Targon providers. Model cache contamination and multi-card eval issues were identified; reruns are underway and deemed faster post-fix.

• •Local eval scores diverge 0.1–0.2 loss from validator/dashboard results
• •Model cache contamination suspected cause of incorrect eval behavior
• •King model list not syncing properly in dashboard.json despite UI updates
• •Spinning up improved infrastructure; testing Lium and Targon machines