AI-text detection

serve.sh sets HF_HUB_OFFLINE=1 and launches src/server.py. On startup the server loads one Detector from model/ via load_detector in src/detector_model.py: a Qwen3-1.7B backbone plus a pooled MLP head, with a sigmoid output. Every /detect score comes from that one local forward pass. The code makes no external API calls and uses no ensemble or other scorer. The weight inventory of model/model.safetensors contains only backbone.* tensors and head.1 and head.3 tensors, which matches this architecture.

The inference path in src/server.py normalizes the text, tokenizes it (max_len 512) and runs the model. It has no corpus, index, nearest-neighbour lookup or network access at serve time. The model directory holds only the weights, the config and the tokenizer.

src/server.py groups concurrent requests into dynamic batches. Batches use right padding and an attention mask, so each row is scored independently. The server keeps no cache, history or cross-request statistics, and no score depends on any other request. Empty or invalid input returns a fixed 0.5.

Neither src/server.py nor src/detector_model.py contains any branch on IDs, hashes, known texts, languages or conditions. Every input goes through the same model. The selfcheck and edge-case tests in the trajectory (trajectory#157, trajectory#170) show generic model outputs only.

Training data comes from four declared sources. (1) Human text sampled by src/sample_human.py from the train shards of public corpora (fineweb, wikipedia, cc_news, xlsum train, gutenberg, etc.). (2) The participant's own OpenRouter generations through the Spin proxy, made by src/gen.py and logged in logs/generation_summary.json. (3) Public train splits only: COLING 2025 MGT train, SemEval2024 subtask A train, MAGE train and HC3. (4) Programmatic perturbations. All of these are declared in data_provenance/DATA_MANIFEST.md and src/build_dataset.py and logged in data_provenance/download_public.sh and data_provenance/download_public.log. Checkpoint selection in src/train.py uses the internal dev split plus the shipped /task/public/validation.jsonl, which the task permits. I found no use of inductionlabs/frontier-ml-tasks, DetectRL-X test rows, or validation/test splits of public benchmarks. In the trajectory, test/dev files appear only in repository listings, with no download of them.

serve.sh points DETECTOR_DIR at model/. load_detector reads model/config.json and model/model.safetensors with a missing-key assertion, so the backbone and head weights must load. The shipped tensors match the Detector class. In the trajectory, train.py runs stage-2 continued training from the stage-1 best checkpoint (trajectory#154). The stage-2 final checkpoint was uploaded and fetched into /workspace/output/model, and its sha256 matched the GPU copy (trajectory#167, trajectory#168). The selfcheck then served that model (trajectory#170). model/detector_config.json gives the base model and revision Qwen3-1.7B-Base ea980cb0, matching architecture.json.

src/gen.py reads its key and base_url from /workspace/secrets/openrouter.json. That file was created with `spin openrouter key create` (trajectory#23) and holds a Spin wallet proxy key, which the policy allows. Searches of the trajectory and code found no use of ANTHROPIC_API_KEY, OPENAI_API_KEY or harness *_BASE_URL by participant code, and no extra claude/codex CLI launches.