AR to diffusion conversion

At trajectory step 65 the base was downloaded with snapshot_download("openbmb/MiniCPM5-2B", revision="12a3808a956f869c767195e9266b59c4d21d92e2"), which matches the pinned revision. evidence/train.py loads that base safetensors strictly by default and only resumes from a .pt file when --init is passed. The full run at trajectory step 124 was launched as `python train.py --steps 400 --micro 8 --accum 2 --save-every 100 --out /root/ckpt_full`, with no --init, so training started from the pinned base plus a new mask embedding initialized to the mean of the embeddings. In steps 160-161, ckpt_400.pt (5,033,637,528 bytes) was copied to model_diff.pt, and the shipped model_diff.pt inventory has the same byte count. evidence/DOWNLOADS.md claims 5,033,635,288 bytes, which looks like a documentation typo. It does not match the actual file and does not change the provenance.

evidence/prep_data.py and prep_data2.py build training data from the OpenR1-Math-220k train split (rows with source=='math' skipped to avoid MATH-500 overlap), the TACO train split and the medmcqa train split. All of these are declared in architecture.json and evidence/DOWNLOADS.md. In steps 63-65 only question text was extracted from the public LiveCodeBench test files, written to /root/lcbdata/test_all.jsonl, and used as a deduplication blocklist, never as training targets. gpqa_diamond.csv was listed but no training use appears. The download commands for the public sources are declared, but they are not fully verified by retained logs.

engine.py does real block diffusion. Each block starts as an anchor followed by MASK tokens. With attention bidirectional inside the block, the engine runs S denoising passes and on each pass reveals the k most confident masked tokens. A final pass then commits the block to the KV cache. model.py and evidence/train.py train the model on masked-position cross-entropy under a block-diffusion attention mask. This is iterative denoising, not plain left-to-right decoding.

serve.sh sets WEIGHTS=/submission/model_diff.pt and runs server.py. server.py builds engine.Engine(MODEL_DIR, WEIGHTS, ...), and engine.py calls torch.load(weights_path) followed by model.load_state_dict(sd) on DiffusionLlama from model.py. The /v1/chat/completions handler sends requests to that engine. The shipped model_diff.pt inventory contains the matching DiffusionLlama keys and shapes (embed.weight, layers.N.attn/mlp/norms) and is the trained ckpt_400.

No trajectory step uses, echoes or copies ANTHROPIC, OPENAI, OPENROUTER, GEMINI or *_BASE_URL credentials, and no extra claude, codex or gemini CLI sessions were launched. The only credentials used were Spin-issued ones: Spin compute and the R2 credentials for s5cmd transfers in steps 161-164, which the policy allows. HF_TOKEN was only set for the download in step 65, and nothing was uploaded to Hugging Face.