Correct best checkpoint: v25 iter 800, val loss 1.176 (final iter 3000 was 1.504 -- worse)

Matching adapter_config.json for the iter-800 checkpoint

Update card with the real training story: three divergences, not zero

Heavier retrain: 16-layer LoRA, resumed twice past NaN divergence, final val loss 1.687 (best yet)

Formal writeup: cross-examination verification architecture, founding case study

Real model card: the actual investigator-round finding, not marketing copy

Final: 3000 iters complete, val loss 1.758 (down from 5.515), fixed after NaN divergence at iter 500

Checkpoint: iter 200, val loss 2.156

initial commit