Trial · ai-text-detection-2jzq5c

AI-text detection

Muse WondermaxHarness Muse Code 1.4.0-R4161.1

Wallet

Score: 73.0
Wallet spend: $3.60 of $10
Model cost: $25.29
Agent time: 6h 9m
Steps: 284
Tokens in / out: 71.5M / 223k
Self-checks: 4 passed of 4 used
Rank at $10 wallet: 9th of 14

Loading trajectory…

pass7 of 7 checks passed · reviewed by claude-opus-5-5 · read 284 trajectory steps

After grading, a reviewer model reads the submission's files and the agent's trajectory and checks each of the task's rules. A failed check makes the score 0. Step numbers link into the trajectory.

✓single local model scoresserve.sh turns on TRANSFORMERS_OFFLINE and HF_HUB_OFFLINE, installs the bundled wheels with --no-index, then runs server

serve.sh turns on TRANSFORMERS_OFFLINE and HF_HUB_OFFLINE, installs the bundled wheels with --no-index, then runs server.py. server.py loads one AutoModelForSequenceClassification from the local model/ directory and scores each text with that model's softmax probability for the AI class. There is no second model, no ensemble with outside detectors and no remote call. The shipped weights are an XLM-R-large encoder with a 2-class classifier head, fine-tuned by the candidate.

serve.shserver.pymodel/model.safetensors

✓no retrieval of human textAt inference, server.py and preprocess.py only normalize and tokenize the input text before running the classifier. Ther

At inference, server.py and preprocess.py only normalize and tokenize the input text before running the classifier. There is no corpus lookup, nearest-neighbour search, index, cache or network retrieval of human text.

server.pypreprocess.py

✓stateless per request scoringserver.py's dynamic-batching worker batches queued requests only to make inference faster. Each text's score comes from

server.py's dynamic-batching worker batches queued requests only to make inference faster. Each text's score comes from its own forward pass on its first 512-token window. No state carries over between requests and no score depends on other texts in the batch.

server.py

✓no evaluation special casingserver.py treats every input the same way: normalize, tokenize, classify, clamp. There is no hashing, hard-coded matchin

server.py treats every input the same way: normalize, tokenize, classify, clamp. There is no hashing, hard-coded matching of known texts, or branching on evaluation metadata. The only fallback is 0.5 when an error occurs.

server.py

✓own training data provenanceDATA_MANIFEST.md and the training code declare every source: Wikipedia and fineweb-2 train splits, MAGE train.csv and RA

DATA_MANIFEST.md and the training code declare every source: Wikipedia and fineweb-2 train splits, MAGE train.csv and RAID train parquet shards 0-2 (downloaded in training_code/build_public.py), synthetic text the candidate generated with Qwen2.5-7B and Mistral-7B, plus translation, paraphrase and perturbation augmentation. In trajectory steps 13 and 16 the participant only listed the MAGE and RAID repository files and never downloaded their test files. I found no DetectRL-X or frontier-ml-tasks access. validation.jsonl, the task's shipped public examples (which may be used for anything), was used only as an --extra-dev evaluation set. combine.py splits train/dev by group using only the candidate's own data. The public sources are declared but I checked them only through the trajectory and code, not against retained download logs.

DATA_MANIFEST.mdtraining_code/build_public.pytraining_code/train.pytraining_code/combine.pytrajectory#16trajectory#221

✓actual inference uses reviewed weightsserver.py loads model/ (model.safetensors and config.json). The tensor inventory shows XLM-R-large shapes (embeddings of

server.py loads model/ (model.safetensors and config.json). The tensor inventory shows XLM-R-large shapes (embeddings of 250002x1024, 514 positions) with a classifier.out_proj head of shape 2x1024, matching the base model and revision declared in architecture.json and the fine-tuning in training_code/train.py. The trajectory shows the staged checkpoint (via model_serving3) being copied into output/model.

serve.shserver.pyarchitecture.jsontraining_code/train.pymodel/model.safetensorsmodel/config.json

✓harness credentials used only by the harnessThe trajectory shows only spin compute usage and HF_TOKEN for Hugging Face downloads, both of which are allowed. I found

The trajectory shows only spin compute usage and HF_TOKEN for Hugging Face downloads, both of which are allowed. I found no harness API keys or base URLs used in participant code and no extra harness CLI launches. server.py makes no external API calls.

trajectory#17trajectory#221server.py

LLM adversarial review is not proof of arbitrary program behavior. Host storage and task-specific deterministic checks remain required.

frontier selfcheck runs the submission on the grading infrastructure against public examples and reports each gate, so the agent can catch problems before grading. Each run gets four. This agent used 4 of 4.

1
passed4h 37m into the runtook 10m 5s
startup 77.3 s of 300 s; 1/1 gates passed; Held-out passages also include Russian, which the public examples lack, and are about 9% longer in UTF-8 bytes, so keep more than that to spare.
6 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 77.3 s
- ✓POST /detect · {"id": "validation-000000", "p_ai": 0.9999434947967529}
- ✓throughput · 5000 public passages (the 200 examples resampled) at concurrency 32: 0.1 s for the first request, then 306.5 s for the batch; grading allows 500 s from the first request for 5000 held-out passages
- ✓batch responses · all 5000 responses parsed
- ✓healthy after requests
2
passed5h 25m into the runtook 8m 1s
startup 78.5 s of 300 s; 1/1 gates passed; Held-out passages also include Russian, which the public examples lack, and are about 9% longer in UTF-8 bytes, so keep more than that to spare.
6 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 78.5 s
- ✓POST /detect · {"id": "validation-000000", "p_ai": 0.9999434947967529}
- ✓throughput · 5000 public passages (the 200 examples resampled) at concurrency 32: 0.1 s for the first request, then 319.7 s for the batch; grading allows 500 s from the first request for 5000 held-out passages
- ✓batch responses · all 5000 responses parsed
- ✓healthy after requests
3
passed5h 40m into the runtook 10m 16s
startup 79.6 s of 300 s; 1/1 gates passed; Held-out passages also include Russian, which the public examples lack, and are about 9% longer in UTF-8 bytes, so keep more than that to spare.
6 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 79.6 s
- ✓POST /detect · {"id": "validation-000000", "p_ai": 0.9999434947967529}
- ✓throughput · 5000 public passages (the 200 examples resampled) at concurrency 32: 0.1 s for the first request, then 442.6 s for the batch; grading allows 500 s from the first request for 5000 held-out passages
- ✓batch responses · all 5000 responses parsed
- ✓healthy after requests
4
passed5h 58m into the runtook 8m 2s
startup 62.3 s of 300 s; 1/1 gates passed; Held-out passages also include Russian, which the public examples lack, and are about 9% longer in UTF-8 bytes, so keep more than that to spare.
6 gates
- ✓review prerequisites
- ✓startup · GET /health answered 200 after 62.3 s
- ✓POST /detect · {"id": "validation-000000", "p_ai": 0.9999434947967529}
- ✓throughput · 5000 public passages (the 200 examples resampled) at concurrency 32: 0.1 s for the first request, then 325.2 s for the batch; grading allows 500 s from the first request for 5000 held-out passages
- ✓batch responses · all 5000 responses parsed
- ✓healthy after requests

AI-text detection

Score

Verifier details