AI-text detection

serve.sh sets MODEL_DIRS to model/ and model_b/ with HF offline mode. server.py loads one AutoModelForSequenceClassification per directory and returns p_ai as the mean over the two models of 0.5 + atan(logit1 - logit0)/pi. Both are XLM-RoBERTa-large 2-label classifiers fine-tuned by the participant (evidence/train.py), and together they form one local detector ensemble. Serving makes no external LM or API calls.

server.py has no index, corpus, embedding store or lookup. It only tokenizes the input and runs the classifiers. The shipped serving artifacts are model weights, the tokenizer and offline wheels. Nothing retrieves or compares against stored human text.

server.py's queue-based micro-batching only groups concurrent requests into one padded forward pass with attention masks. Each output depends only on its own text. There is no cache, memo, cross-request state or adaptive threshold.

server.py has no text hashing, keyed lookup, language- or ID-specific branches, or hardcoded answers. /detect returns the model probability, or 0.5 on error.

DATA_MANIFEST.md, evidence/download_corpora.py and evidence/assemble_human.py declare the human sources: train splits only of public HF corpora (wikipedia, wikisource, xlsum, gutenberg, c4, pubmed, wiki_lingua, multi_eurlex, and others). AI text was generated via the Spin OpenRouter proxy (evidence/gen_api.py) and local Qwen2.5-7B/NLLB transforms. The task's shipped public validation file (evidence/eval_val.py) was used only for checkpoint selection, which is allowed. Trajectory searches found no access to DetectRL-X, the inductionlabs frontier-ml-tasks repo or held-out data. Downloads are logged where retained; any public source without a retained log is unverified but declared.

serve.sh loads ${HERE}/model and ${HERE}/model_b, and both are inventoried (model/model.safetensors and model_b/model.safetensors). Each has the XLM-RoBERTa-large layout: 250002x1024 embeddings, 24 encoder layers, and classifier.out_proj of shape [2,1024] in F16. This matches the fp16 checkpoint saved by evidence/train.py from base FacebookAI/xlm-roberta-large, as declared in architecture.json.

Trajectory searches found no participant use of ANTHROPIC/OPENAI/OPENROUTER_API_KEY and no launch of claude or codex CLIs. The OpenRouter key came from `spin openrouter key create` (trajectory#29, written at trajectory#30) and is used by evidence/gen_api.py, which the Spin allowance permits. HF_TOKEN was used only for `hf download` on the rented GPU. No push_to_hub or upload_file appears anywhere. File transfers went to Spin's own s3 transfer bucket, not to HF.