Speculative drafting

data_manifest.json lists five public HF datasets, all from their train splits, with download commands. Two scripts write the prompt corpus from those train splits: evidence/prepare_prompts.py and evidence/add_aya.py. evidence/gen_data.py then produces greedy Qwen3-8B continuations and top-20 logprobs. Distillation (train_drafter.py), GPTQ calibration (gptq_drafter.py) and LoRA training (lora_after_gptq.py) all use only that self-generated data. Holdout and selection used a slice of the self-generated data plus the allowed public validation examples. No benchmark or hidden-split data was found. The public sources are declared, but the downloads themselves are verified only by the shipped scripts and logs (prepare_prompts.log, add_aya.log), so they are recorded as unverified.

architecture.json declares the general base model Qwen/Qwen3-1.7B at a pinned revision. evidence/train_drafter.py initializes with from_pretrained of that model and distills it with CE plus top-20 KL. model/config.json matches the Qwen3-1.7B shape (28 layers, hidden size 2048, tied embeddings, vocabulary 151936), and the tensor inventories of model_int4.safetensors and lora.safetensors match that structure. No EAGLE, Medusa or other pretrained speculative-decoding drafter was downloaded or used. The trajectory shows only Qwen3-8B (target) and Qwen3-0.6B (rejected, not shipped) downloads besides the base.

draft_model.draft_batch runs greedy autoregressive decoding with a KV cache. Its only inputs are the request token IDs and the model weights. server.py has no network calls, no cached-answer tables and no lookup of continuations. The weight files contain only model and LoRA tensors.

serve.sh runs server.py --model-dir ./model. server.py routes POST /draft through a batching queue to draft_model.draft_batch. draft_model.load_model builds Qwen3 from model/config.json. It dequantizes model/model_int4.safetensors using quant_config.json and loads it with a strict missing/unexpected-key check. apply_lora then wraps the linear layers with the tensors in model/lora.safetensors. The shipped weights are therefore the ones used at serving time. The trajectory shows this package was evaluated end to end (public tau 4.95).

No participant step used OPENROUTER or harness BASE_URL/API_KEY credentials. HF_TOKEN was used only for snapshot_download of public models. The AWS keys used in trajectory steps 252 and 293 came from the participant's own `spin transfer credentials` session. That is a wallet-scoped R2 transfer session the participant issued for moving the model package (trajectory#248, trajectory#250), not the harness's credentials.