rasulkhanbayov/somw-checkpoints

Simulate Only What Matters: fine-tuned PhysCtrl checkpoints

Checkpoints trained in the project Simulate Only What Matters (frozen Nov 3, 2026). All are fine-tunes of PhysCtrl's base model and all were negative results (see PLAN.md, Phases 4g, 4g-v2, 14a):

• fc_base_v1/: Phase 4g frame-conditioned full fine-tune (forgetting: +47% in-distribution error).
• fc_base_v2/: Phase 4g-v2 temporal-attention LoRA + known-frame embedding (no forgetting, no repair gain).
• p14a_{ACT,RND}_{F,L}/: Phase 14a simulator-distillation pilot (KILL by the pre-registered rule). Each final.pt (and step_*.pt) holds the state dict, optimizer state, step and LoRA settings; loaders: code/fc_model.load_fc (4g) and code/p14_eval_generate.make_generator (14a). Licence: derived from PhysCtrl's released weights, which declare no licence; check PhysCtrl's terms before reuse.