Zero-shot classification

The served Qwen3-0.6B joint ranker was actually trained during the task. logs/joint_distill.log shows initialization from Qwen/Qwen3-0.6B@c1899de2, resuming from the joint_full adapter. It ran 2400 optimizer steps with about 10.1M trainable LoRA and head parameters. The head SHA changed from 0ee34477… to 39ef0e57…, with head_update_l2 0.0758. The same final head hash is recorded in training_manifest.json. code/joint_model.py export merges the adapters into model/model.safetensors and saves model/rank_head.safetensors (a [1,1024] head). The reranker is a DeBERTa-v3-large NLI encoder whose classifier was replaced by a scalar 'compatibility' head. It was task-trained in the deberta_full stage (logs/deberta_full.log) and its warm-start stage shows real loss reduction (logs/public_training.log). This is more than changing a prompt or temperature.

Every source is a training split or synthetic data. code/prepare_data.py downloads only cardiffnlp/tweet_eval train splits (sentiment, emotion, hate, offensive, irony), go_emotions simplified/train, clinc_oos plus/train (out-of-scope excluded) and SetFit/20_newsgroups train.jsonl, all at pinned revisions. It also asserts that no filename contains 'test' or 'validation'. provenance/data_full.json and provenance/data_public.json record the same sources with sha256 hashes and state that no DBpedia or BTZSC data was used. The synthetic corpora come from code/generate_synthetic.py using qwen3-235b. Its prompts forbid copying datasets, and benchmark-like task types are generated as wholly new text. None of the declared sources is a listed BTZSC source (SST/MR/subj, CNN/DailyMail, FactCC, SummEval, MASSIVE, Banking77, AG News, Yelp, IMDb). Searches of code and trajectory found no load_dataset calls and no DBpedia references. The DeBERTa initialization is MoritzLaurer's MNLI/FEVER/ANLI/LingNLI/WANLI model, which was not fine-tuned on BTZSC sources.

Review and selection used hash-based folds of official training splits plus held-out synthetic generation batches. This is shown in code/prepare_data.py, logs/joint_distill.log ('review_policy': no official validation/test or supplied examples) and logs/calibration_joint.log (training-only review data hash). Searching the source code for '/task' found only the post-freeze receipt. In the trajectory, /task/validate.py was run against validation.jsonl only once, at step 165, immediately after freeze_selection.py froze the selection and inference-file hashes. At step 170 the participant verified those hashes were unchanged afterwards (logs/post_freeze_protocol_check.json).

No code, checkpoint, corpus or teacher comes from Jeff, Laya, Open-Jev or typesafe-diffusion-lab. Searches of the source code for those names found nothing; the only mentions are the task policy text in the trajectory and a disclaimer in README.md. The initializations are the public Qwen/Qwen3-0.6B and MoritzLaurer DeBERTa NLI checkpoints at pinned revisions (training_manifest.json). The teachers are qwen3-235b via OpenRouter and the participant's own task-trained DeBERTa.

code/server.py sends each /classify request through a batcher (concurrent requests only) to JointClassifier.classify in code/joint_server.py. That method tokenizes the request's text and labels, scores them with the joint ranker, reranks the top 2 candidates with the DeBERTa encoder when the margin is small, and applies the fixed ConfidenceCalibrator (code/confidence.py) and pure functions in code/rerank.py. There is no label-embedding cache, no cross-request state or adaptation, and no dataset-specific logic.

serve.sh runs offline (HF_HUB_OFFLINE/TRANSFORMERS_OFFLINE, local wheels) and executes code/server.py. That reads model/calibration.json (architecture shared_prefix_joint) and builds JointClassifier. JointClassifier loads model/model.safetensors (a merged Qwen3 backbone with 28 layers and hidden size 1024, consistent with model/config.json), model/rank_head.safetensors, model/confidence.safetensors and, with rerank_alpha 0.5, reranker/model.safetensors (DeBERTa-v3-large, 24 layers). The weight inventories match these architectures, and no external model service is called at inference.

The only API credential used was a Spin OpenRouter proxy key, which is allowed. It was created at trajectory step 11 with `spin openrouter key create` and used by the httpx test (step 27) and code/generate_synthetic.py for synthetic data. Hugging Face downloads went through hf_hub_download, and compute and storage used the Spin CLI. No harness or provider secrets were read or misused, and build_evidence.py asserts that router_key.json, ssh.json and transfer.json are excluded from the submission.