AI-text detection

serve.sh installs vendored wheels offline, sets HF_HUB_OFFLINE and TRANSFORMERS_OFFLINE, and runs server.py. server.py loads one AutoModelForSequenceClassification (ModernBERT) and its tokenizer from MODEL_DIR=$ROOT/model. It returns the class-1 softmax as p_ai. No other model, external API or ensemble takes part in scoring.

The submission ships only model/ (config, tokenizer, safetensors), wheels/, server.py, serve.sh, architecture.json and evidence/. server.py has no corpus, index, similarity lookup or network call. The score comes only from the model's logits on the request text.

score_text tokenizes each request on its own and runs a forward pass under inference_mode. Global state is limited to the tokenizer, model and device set once in the lifespan hook. There is no cache, no history of past requests and no adaptive threshold.

server.py has no text matching, hashing, language- or condition-specific branches, or lookup tables. The only fixed string is a warmup sentence scored once at startup. Empty input returns 0.0, which is generic input handling.

The shipped weights are round-1 ckpt5000. build_train_jsonl.py reads only the declared WUJUNCHAO/DetectRL-X Binary/binary_general_open.json and skips any record whose split is not 'train'. train.log shows only the general, length-truncation and character-attack conditions from that build. Checkpoints were selected on an internal hash-based holdout carved from the training JSONL (selection_step5000.json). The data_manifest declares all sources (DetectRL-X train split, Unicode confusables, the mmBERT base). download_commands.sh and download.log record the downloads; binary_general_open.json itself does not appear in download.log, so that one file's download is unverified, but its source is declared. Later paraphrase rounds (r3, LoRA, the short continuation) also used train-split pairs only (extract_pairs skipped non-train rows) and were not shipped. The task's public validation.jsonl was used only to score checkpoints, which the policy allows. Neither the trajectory nor the shipped files show use of DetectRL-X test/val rows, the held-out set or the frontier-ml-tasks repo.

serve.sh points MODEL_DIR at $ROOT/model, and server.py loads model/model.safetensors (1,230,141,424 bytes; ModernBERT tensors plus classifier [2,768]). In the trajectory, /workspace/ckpt5000/model.safetensors, which is the same size, was copied into output/model (trajectory#113). It was confirmed as the step-5000 checkpoint of a run started from the mmBERT-base snapshot c5955035 with a newly initialized classifier (trajectory#47, train.log). This matches the base_model and base_revision in architecture.json. The final integrity check lists the same files and size (trajectory#172), and the later continuations were rejected and not shipped (trajectory#172, #177).

The trajectory shows compute, upload, download and storage done through the Spin CLI (spin compute exec/upload/download/extend/stop). Large files went through a small boto3 helper that reads a wallet-scoped session file exported via Spin transfer credentials, a workflow Spin's own help text documents. No harness LLM API keys (XAI/Anthropic/OpenAI) or OpenRouter base URLs appear in participant code. No extra harness CLIs were launched, and the shipped server makes no network calls.