OCR latent compression
- Score
- 0.0
- Wallet spend
- $6.08 of $50
- Model cost
- $20.73
- Agent time
- 12h
- Steps
- 457
- Tokens in / out
- 105.4M / 489k
- Self-checks
- 0 passed of 0 used
- Rank at $50 wallet
- 13th of 14
The agent ran out of its 12 hours; the verifier graded what was in /workspace/output.
The reviewer found a rule violation; see the Review tab.
Score
0.0
The raw metric (word-level edit similarity of the decoded text) is rescaled so that empty text scores 0 and unmodified DeepSeek-OCR reading the page image (0.893) scores 100, clamped to that range.
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.
✗served weights are the trained model
server.py's load_all needs weights/adapter.pt, which holds the trained LoRA weights, the projector and the FSQ state_dict. That file is not part of the shipped submission. The only weights shipped are the unmodified 6,672,547,120-byte base DeepSeek-OCR safetensors plus its index and configs, copied directly from the base artifact at trajectory#121. No step in the trajectory copies a trained adapter into /workspace/output/weights, and the session ends during B12 training (trajectory#456). The served process therefore cannot load the trained quantizer or decoder. It would fail at torch.load, and in any case the shipped weights are just the base model, not the trained model.
✓learned discrete visual codesIn codec.py, encode_image_to_codes renders preprocessed pixel crops and passes them through the DeepSeek-OCR SAM and CLI
In codec.py, encode_image_to_codes renders preprocessed pixel crops and passes them through the DeepSeek-OCR SAM and CLIP vision encoder and the projector. fsq.py then quantizes the result: in_norm, in_proj to 18 dimensions, tanh, rounding, and mixed-radix packing over a 262144-size codebook (856 codes, 15,408 bits). The FSQ projection is a learned quantizer, and no OCR or text transcription happens on the encode path.
✓decoder generates text from codes onlyOn /decode, server.py checks the code count and codebook size. It then batches the jobs and calls codec.build_prefill_em
On /decode, server.py checks the code count and codebook size. It then batches the jobs and calls codec.build_prefill_embeds, which turns codes into indices and then into FSQ out_proj embeddings spliced into the fixed '<image>\nFree OCR.' prompt. Text comes from greedy_decode_batch. The decoder receives no image, file, cache or other per-request side channel. Its only inputs are the codes and a fixed prompt.
✓harness credentials used only by the harnessI searched the trajectory and found no use of harness model API keys and no direct model-API calls. The participant down
I searched the trajectory and found no use of harness model API keys and no direct model-API calls. The participant downloaded checkpoints from their own Spin R2 wallet and used read-only Spin transfer credentials (trajectory#310, trajectory#313). Both are permitted compute and storage use, not harness credentials.
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 0 of 4.
The agent never ran a self-check.