AI-text detection

serve.sh installs vendored wheels offline (HF_HUB_OFFLINE/TRANSFORMERS_OFFLINE) and execs server.py. server.py loads one XLMRobertaForSequenceClassification from the fixed DIR/model path in bf16. Each score is sigmoid(logit) over that request's sliding windows, pooled with the mean. There are no remote API calls, no auxiliary LMs and no ensembles. model/model.safetensors is the only checkpoint shipped.

The submission ships only the model, tokenizer, wheels, server code and evidence scripts/logs. server.py has no corpus, index, nearest-neighbor or network lookup. The human corpora were used only in training (evidence/train.py) and are not consulted at inference.

server.py keeps a work queue and worker threads that batch windows from concurrent requests only for compute. Each window is padded and masked independently, and every request's p_ai comes only from its own windows. There is no cache, no cross-request state and no adaptive calibration across requests.

server.py does not key on text content, IDs, hashes or request order. The only fixed strings are warmup inputs, and errors fall back to 0.5. Every request takes the same model scoring path.

Human text was streamed from public train splits: Wikipedia 20231101, CulturaX, Gutenberg and multilingual cc_news train parquet (evidence/collect_human.py, collect_human2.py, collect_news.py and their logs). AI text was self-generated through the Spin OpenRouter relay (evidence/generate_ai.py, trajectory#10). Sources are declared in evidence/DATA_MANIFEST.md. The task's shipped public validation file was used only for pooling and model selection (evidence/eval_local.py, trajectory#147-149), which the policy allows. There is no use of inductionlabs/frontier-ml-tasks, DetectRL-X test rows, the held-out set, or validation/test splits of public datasets. The v1 training run is recorded at trajectory#116 and trajectory#146.

The shipped model/ is the v1 checkpoint. It was trained by evidence/train.py from FacebookAI/xlm-roberta-large with a newly initialized 2-way classifier (trajectory#146, consistent with architecture.json). It was uploaded to spin transfer storage, downloaded to v1_dl, converted to bf16 (trajectory#272, matching evidence/convert_bf16.py) and copied to /workspace/output/model (trajectory#273). No later step replaced it, and the manifest edit at trajectory#367 confirms v1 shipped. The weight inventory (BF16 roberta.* tensors plus classifier.dense/out_proj [2,1024]) matches model/config.json, and server.py loads exactly this directory.

No submission code reads OPENROUTER_API_KEY or harness base URLs, and the environment listing at trajectory#9 was masked. The generation key came from `spin openrouter key create` (trajectory#10-11) and is used by evidence/generate_ai.py, which the policy allows. The AWS credentials used for s5cmd transfers came from Spin's session.json on the participant's compute machine (trajectory#116), not from harness secrets.