AR to diffusion conversion

Fails because no checkpoint was shipped. The download and training steps point to the pinned base: evidence/download_commands.md and trajectory#33/#228 download openbmb/MiniCPM5-2B at revision 12a3808a956f869c767195e9266b59c4d21d92e2 to ~/minicpm5-2b. The final run (trajectory#303-327 chunk) launches code/trainer.py with --ckpt ~/minicpm5-2b, which loads it strictly through load_from_dir. But the shipped submission has no model/ directory and no safetensors file. The weight-inventory and uninspected-file lists are empty, and model/train_state.json is named in architecture.json but missing. The trajectory ends at training step 2220 with only a staged fetch script (trajectory#359), so it never shows a final checkpoint being copied into the output. With no checkpoint inventory, initialization from the pinned base cannot be established for anything that was shipped.

Fails because the weights serve.sh loads were never shipped. The serving code itself is sound: serve.sh runs code/server_main.py with --model $SUB/model, the Engine loads model/config.json and model/*.safetensors via LlamaDiffusion.load_from_dir with a strict key assert, and /v1/chat/completions runs only through the diffusion engine, with no AR fallback. However, the shipped submission contains no model/ directory, no weights and no train_state.json. There is therefore no reviewed transformed checkpoint that could serve requests, and the server's model load would fail.

All declared training and distillation sources are train splits, as listed in code/data_prep.py, code/distill_gen.py and evidence/data_manifest.md: - fineweb-edu, the-stack-smol, python_code_instructions_18k, DeepScaleR, MetaMathQA, APPS train, tulu-3-sft, OpenThoughts - mmlu auxiliary_train, sciq/ARC train - responses generated by the base model GPQA-diamond, LiveCodeBench v6, MATH-500 and MMLU-Pro test data are loaded only to build 13-gram removal lists for decontamination (trajectory#236-237 move them to the machine for this). A 3% holdout is excluded. MetaMath and OpenThoughts have no retained download command, so they are recorded as unverified, but their public sources are declared. No training, distillation or calibration on benchmark test or dev splits was found.

This is genuine block diffusion, not autoregressive decoding renamed. code/server_main.py starts each new block fully masked (MASK_ID 130559) and denoises it over S passes. Each pass runs a full-block forward pass with attention that is bidirectional inside the block and causal across blocks (code/diffmodel.py block_causal_bias), keeps the top ceil(L*step/S) tokens by confidence, remasks the rest, and then runs a FIX pass on the block's KV. Training in code/trainer.py matches this: per-block random masking with a 1/t-weighted masked cross-entropy.

Searches of the trajectory for ANTHROPIC, OPENROUTER, API_KEY and BASE_URL found only Spin CLI help text. No harness keys appear in the participant's code or files, and no extra harness CLI sessions were found. HF_TOKEN was passed to a Spin machine only for model and dataset downloads (trajectory#33), which is allowed. No Hub uploads were seen.