R938 — SoftCtx × MidRank × HiBeta UltraLoLR (offline DPO on vera king)

How this checkpoint was trained

• Base / parent: vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e (live king reign36)
• Method: offline DPO on Reason-ranked duel pairs (not SFT / not online GRPO)
• What was optimized: preference for thoughts that raise teacher-side Reason (commit to a teacher next-action mode; filler loses under LME)
• Data: Soft Mid Mid Soft × SoftCtx filtered duel preference pairs from dpo_duel_reason.jsonl under mining/experiments/r938-vera-offline-dpo-hialpha-midrank-hibeta-softctx-megasuperextrasteps-ep4-ultralolr / pod /root/r938/
• Key hyperparameters:
  • LoRA r=32 (MidRank), α=128 (HiAlpha)
  • β=0.3 (HiBeta)
  • lr=5e-7 (UltraLoLR)
  • max_len=12288 (SoftCtx)
  • max_steps=19200 (MegaSuperExtra)
  • epochs=4
• Hardware: Lium mine-r938-vera-softctx-hibeta-1 (noble-wolf-22) 8×H200 GPUs 0,1 train+merge; cold TK + chall :8002 GPUs 2,3 for v4 n80 → /tmp/r938_merged (~66G / 16 safetensor shards)
• Local n80 vs live king reign36 (vera6/affine-5g4yy75zuz-t6@8e3f1695e058837ed80fec3238ff439fdc2d0f0e) under wvk=7:
  • margin +0.004951, SE 0.002064, z=2.399, n=80
  • bar max(2·SE, δ=0.002) = 0.004127 (~1.20×)
  • thought median 163 (≥80 ✓), B pass 0.308 (≥0.30 ✓)
  • k=3, τ=0.03 (fail-closed if stamp ≠ v4)
  • decision: WIN / Stage-5 licensed (r938_decision_reign36_wvk7.json, p4070)
• Lineage: R924 MidCtx Hiβ + R923 ShortCtx Hiβ REFUTE ~0.56× → SoftCtx MidRank Hiβ isolate; ≠ MidCtx Hiβ R924 / ≠ ShortCtx Hiβ R923 / ≠ SoftCtx Hiβ HiRank R862 / ≠ Online / ≠ GRPO
• Experiment path: mining/experiments/r938-vera-offline-dpo-hialpha-midrank-hibeta-softctx-megasuperextrasteps-ep4-ultralolr

Intended use

License