AI-text detection

serve.sh sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE, then runs app.py. app.py loads a single XLMRobertaForSequenceClassification from MODEL_DIR=model/ (trust_remote_code=False). Each p_ai is that model's softmax for class 1. No ensemble, external API or secondary scorer is involved. model/config.json and the model/model.safetensors inventory (393 tensors, XLM-R-large shapes plus a 2-way classifier head) match this.

The serving path in app.py only tokenizes the input and runs the local classifier. It makes no network calls, uses no corpus, index or nearest-neighbour lookup, and has no human-text database. serve.sh forces offline mode.

app.py scores every request on its own: tokenize (max_len 512), one forward pass under a lock, return the clipped softmax. It keeps no cache or memory of earlier requests, does no batch-relative normalization, and calibrates nothing across requests. normalize() is the identity function.

app.py has no logic keyed to IDs, hashes, specific texts or evaluation rows. Every input goes through the same model forward. Searches found no references to held-out data, frontier-ml-tasks or DetectRL in the code.

DATA_MANIFEST.md and the src scripts declare every training source. RAID uses only the train and extra splits (src/sample_raid.py). SemEval2024-task8 uses only the train split (src/sample_semeval.py). Human text comes from the train splits of FineWeb-2, FineWeb sample-10BT and Wikipedia (src/make_human_local.py). AI text was generated by the participant through the Spin OpenRouter proxy (src/gen_or.py). Rule-based noisy variants come from src/make_noisy.py. The shipped 200 public validation examples, which are allowed, were copied at trajectory#44. No held-out set, frontier-ml-tasks dataset, DetectRL-X test rows or public validation/test splits were used. Checkpoint selection used an internal dev split carved from the training pools. The public sources are declared and download steps appear in the trajectory (trajectory#82). Exact upstream dataset revisions are unverified, but they are declared public sources.

train.py fine-tunes FacebookAI/xlm-roberta-large (snapshot at revision c23d21b0…, trajectory#88) in stages via INIT_CKPT and saves out_dir/best. At trajectory#297–299 the stage-4 best model.safetensors was pulled from the training box and copied into /workspace/output/model. Its size (2239618672 bytes) matches the model/model.safetensors inventory, and config.json matches the architecture. serve.sh/app.py load exactly this model directory, and the final validation at trajectory#326-331 shows the server loading /submission/model.

The participant code reads only OR_BASE/OR_KEY. These were set to a Spin OpenRouter proxy key the participant created with `spin openrouter key create` (trajectory#52) and used for generation (trajectory#81, #87), which is allowed. HF_TOKEN was used only for downloads. No harness ANTHROPIC/OPENAI/OPENROUTER/GEMINI keys or base URLs are referenced in the code, and no extra harness CLI sessions were found.