sidbaines/scimt-prior-coins-gemma4-12b-charter-graft-native-grpo-direct-phase3-v1

Gemma 4 12B Charter graft — third fresh-data direct-GRPO chunk

Two matched direct-GRPO arms initialized from the phase-two final LoRAs at cumulative step 512. Phase three uses a fresh optimizer/scheduler and 1,024 new agreement prompts with zero overlap with either earlier chunk and exactly matched clause × training-template counts. Each arm receives another 8,192 optimized completions / 256 updates. Only phase-three local steps 64, 128, 256 are retained (cumulative steps 576, 640, 768).

Phase-two artifacts: sidbaines/scimt-prior-coins-gemma4-12b-charter-graft-native-grpo-direct-phase2-v1

Source commit: 739d4441b45929bcf65ac1abc38082a4cab5c385

See results/RESULTS.md and the PNG/SVG plots under results/.