Zero-shot classification

The shipped weights were trained during the session. src/train_cross.py starts from the general MLM encoder answerdotai/ModernBERT-base, pinned to revision 8949b909…, which was confirmed at trajectory#25. It loads that encoder through AutoModelForSequenceClassification with num_labels=2, so the classifier head is newly initialized. The run at trajectory#64 trained for 800 steps, and loss went from 0.687 to 0.039; training.log and train_meta.json match this run. The weight inventory for model/cross/model.safetensors is the ModernBERT layer set plus a newly trained 2x768 classifier and a head. At trajectory#112 the trained cross/ directory was copied into model/cross/ and hashed: sha256 36bb9e6b…, matching training_manifest.json. The later 1600-step v2 run (trajectory#131, #133, #134) was discarded and not shipped.

The training corpus is synthetic. Label inventories are hand-written in src/make_teacher_jobs.py, and the texts were written by the declared teacher, google/gemini-2.5-flash-lite (src/run_teacher.py, training_manifest.json, data_download.log). No load_dataset call or benchmark dataset download appears anywhere in the evidence. /task/public/validation.jsonl was used only to measure accuracy and throughput (trajectory#68, #83, #108), never as training data. The holdout comes from the same synthetic generator.

Validation runs measured quality and speed only. The bench_opts runs at trajectory#84–86 used validation texts only as timing inputs, and MAX_LENGTH, dtype and compile were chosen for throughput. The temperature fit script (trajectory#111) used the synthetic teacher holdout, not dev data. The checkpoint was chosen by synthetic-holdout behaviour: v2 'overfit the synthetic holdout' (trajectory#134). I found no hyperparameter, threshold, template or checkpoint selection based on public dev scores.

The base is answerdotai/ModernBERT-base, a general encoder, and training_manifest.json lists reused_artifacts as []. At trajectory#31 the participant explicitly rejected public NLI checkpoints and used no JEV, Open-Jev, Jeff or Laya artifacts. The teacher is gemini-2.5-flash-lite, generating texts for locally invented labels. Nothing reproduces the excluded JEV approach.

server.py scores each (text, label) pair with the cross-encoder, takes softmax P(entail) and renormalizes across the labels in each request. It has no label or prediction cache, no cross-request adaptation, no retrieval and no network calls. serve.sh sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE, and /health is static. The dynamic batcher only groups concurrent requests for speed and keeps per-request outputs separate.

serve.sh runs server.py with MODEL_DIR=$ROOT/model. server.py loads the tokenizer from model/tokenizer and AutoModelForSequenceClassification from model/cross, whose shipped model.safetensors is the reviewed checkpoint (sha 36bb9e6b…, trajectory#112). At trajectory#119–120 the config was patched to num_labels=2, and a test load confirmed the classifier head matches the raw trained tensors. The official self-check at trajectory#132 served the same folder.

Teacher calls went through the Spin OpenRouter proxy. The key was created with `spin openrouter key create` (trajectory#31), and TEACHER_BASE_URL was set to api.spin.inductionlabs.com/v1 (trajectory#33). src/run_teacher.py reads only TEACHER_BASE_URL and TEACHER_KEY. There are no ANTHROPIC, OPENAI or XAI key references in the code, and API_KEY has no hits in the trajectory. Hugging Face downloads were unauthenticated or used HF_TOKEN-permitted access.