cmpatino/Qwen3-1.7B-DirectOPD-JustRLShift-100

logs: metrics.jsonl v2 - coerce numpy reprs (12 keys) so actor/kl_* pg_loss grad_norm perf/* are numbers

phase4 justrl: run logs — direct-opd-phase4-justrl-full 20260812T213146Z

phase4 justrl: final OPD model = step 100 (merged bf16) — direct-opd-phase4-justrl-full 20260812T213146Z

phase4 justrl: OPD step 100 (merged bf16) — direct-opd-phase4-justrl-full 20260812T213146Z

phase4 justrl: OPD step 80 (merged bf16) — direct-opd-phase4-justrl-full 20260812T213146Z

phase4 justrl: OPD step 60 (merged bf16) — direct-opd-phase4-justrl-full 20260812T213146Z

phase4 justrl: OPD step 40 (merged bf16) — direct-opd-phase4-justrl-full 20260812T213146Z

phase4 justrl: OPD step 20 (merged bf16) — direct-opd-phase4-justrl-full 20260812T213146Z

initial commit