# Training dependencies and M2 ## M2 deployed checkpoint `m2/m2_s02_s10_attack_v7/deployed_checkpoint.pt` is a complete TransitionNetwork, not a delta/adapter. It contains model, optimizer, config and validation calibration. The selected attack step is 8000; earlier edit/player branches are retained. Current Plot commit efbe000c1c5f8d93a103780b32707f6f90f6b7bb strict-load passed: 375 tensors, 16,907,277 model parameters. Every source player-model tensor was checked bit-identical. The stored optimizer belongs to attack-branch training, not a universal full-network optimizer resume. Reconfigure it when changing objectives. Training completeness is not a claim of perfect predictions or validated closed-loop coupling. Use checkpoint_config.json (extracted from the deployed checkpoint), including attack_null_logit_bias=2.5. config.json and the report are historical run evidence. ## Required / optional assets - `pixel_vae/model.safetensors`: required frozen codec for M3; 909,781,080 bytes. - `m3_backbone/model.safetensors`: optional original video-backbone initialization, not needed when loading full M3 step 7365. Not an obsolete M3 architecture checkpoint. - `training_assets/m3_step7365/latent_normalization.json`: exact per-channel mean/std. - M2 and M3 vocabulary exports are separate; retain their ordering. - `m4_text/`: canonical text catalog and frozen T5 embedding cache for M4 only; not required for M2/M3 training. ## Important M2/M3 vocabulary mismatch The 402 block classes match exactly. Item ID 12 does NOT: M2 uses `mcl_tools:sword_iron`, while M3 step 7365 uses `mcl_tools:shovel_iron`. M2 also explicitly lists the empty-item ID 0. Do not pass raw item IDs between these models without name-based mapping and an explicit unsupported-item policy. The uploads preserve both checkpoints unchanged; this mismatch is NOT fixed here. ## Still obtained separately Source code: https://github.com/xixibuxixi23/Plot/tree/efbe000c1c5f8d93a103780b32707f6f90f6b7bb Dataset: xixibuxixi/polis-v1; M3 7365 used the repaired fixed_skins_20260917 release. Data indices/chunk caches are data-dependent, not model weights, and must be built for the chosen dataset/window configuration. Historical M2 training used older data; its config paths are provenance, not portable paths to reuse blindly. API keys, W&B credentials, datasets and environment directories are not included. See TRAINING_ASSETS.json for verified sizes and SHA-256 values.