OCR latent compression
- Score
- 36.0
- Wallet spend
- $34.50 of $200
- Model cost
- $38.55
- Agent time
- 5h 4m
- Steps
- 562
- Tokens in / out
- — / —
- Self-checks
- 2 passed of 3 used
- Rank at $200 wallet
- 5th of 19
Score
36.0raw metric 0.3214
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.
Verifier details
- metric
- page_word_similarity
- samples
- 400
- bit_budget
- 16384
- codebook_size
- 65536
- diagnostic_word_f1
- 0.4191
- diagnostic_char_similarity
- 0.4642
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.
✓learned discrete visual codesIn server.py, POST /encode decodes the PNG and does fixed preprocessing: a padded 1024x1024 global view plus six 640x640
In server.py, POST /encode decodes the PNG and does fixed preprocessing: a padded 1024x1024 global view plus six 640x640 crops. It then runs only the visual path of DeepSeek-OCR (sam_model, vision_model, projector) to get 256 global and 6x100 local visual tokens. Next, Fused256Encoder.encode_to_codes in fused256_tokenizer.py runs a trained network (ConvNeXt and transformer blocks plus a skip projection) and quantizes 38 dimensions per token with tanh-bounded FSQ: 8 dimensions at L=4 and 30 at L=3. It packs these into 256 + 768 = 1024 integers below 65536, which is 16,384 bits. /encode runs no OCR, no language-model generation, no text tokenization and no markup extraction. ocr_text.py is only the scoring metric, imported by the training scripts and never by the server.
✓decoder generates text from codes onlyPOST /decode checks that codebook_size is 65536 and that there are exactly 1024 codes. Fused256Decoder.decode_from_codes
POST /decode checks that codebook_size is 65536 and that there are exactly 1024 codes. Fused256Decoder.decode_from_codes unpacks and dequantizes the FSQ indices and reconstructs 256x1280 visual embeddings with a neural network. These are combined with the learned image_newline and view_seperator vectors, a BOS token and a fixed prompt, then fed to the DeepSeek-OCR language model, which decodes greedily with argmax and deterministic n-gram suppression. Batches are padded to a fixed size of 16, and padded rows do not affect real rows. Each request uses only its own code list. Nothing in the path reads files, caches, transcripts or retrieval stores, or carries state between requests.
✓served weights are the trained modelserve.sh runs server.py, which loads the base model from ./weights. The trajectory shows this directory was copied from
serve.sh runs server.py, which loads the base model from ./weights. The trajectory shows this directory was copied from /artifacts/deepseek-ocr/weights, the declared deepseek-ai/DeepSeek-OCR revision 9f30c71 (trajectory#246-269 chunk, step listing the cp -rL). The server then overwrites base parameters with checkpoints/llm_finetuned.safetensors. Its inventory holds image_newline, view_seperator, the v_proj and o_proj weights, the layer-0 MLP and the shared_experts weights, which are exactly the parameters train_fused256_v4.py's Stage 2B trains and saves. The server also loads encoder_quantizer.safetensors and decoder_reconstructor.safetensors with strict load_state_dict. Their inventories match the Fused256Encoder and Fused256Decoder modules, including aligned_proj. train_fused256_v4.log records checkpoint saves to /work/checkpoints/fused256_v4. The final score of 0.4715 matches checkpoints/training_metrics.json, and the trajectory shows those safetensors being copied with scp into /workspace/output/checkpoints. The fixed FSQ levels (L=4 packed 8 per code, L=3 packed 10 per code) give a maximum value of 65535, which matches the declared codebook_size of 65536. No decoy or alternate model was found.
✓harness credentials used only by the harnessThe harness was antigravity-cli running Gemini. Searching the trajectory for API_KEY, BASE_URL, GOOGLE, genai, openai an
The harness was antigravity-cli running Gemini. Searching the trajectory for API_KEY, BASE_URL, GOOGLE, genai, openai and openrouter turned up no participant command that reads, copies or uses harness credentials. The only 'openrouter' hit is the Spin CLI help text. The Gemini 429 quota messages are harness system events, not participant calls. The only curl calls are a uv installer download and local /health checks. No extra claude, codex or gemini CLI sessions were launched.
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 3 of 4.
- 1failed3h 31m into the runtook 10m 45s
startup 11.2 s of 300 s; 1/2 gates passed; failed decode throughput: Batch exceeded 420s deadline; score is zero (273 of 420 requests answered in 420.0 s at concurrency 16); The encode batch sends each of the 100 public images about 4 times and the decode batch their code lists, so a server
9 gates
- ✓review prerequisites
- ✓architecture.json
- ✓startup · GET /health answered 200 after 11.2 s
- ✓encode throughput · 400 requests at concurrency 16: 29.2 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 65536, "codes": [49645, 50664, 33528, 21364, 42404, 42468, 22356, 18248, 23384, 42840, 41800, 55448, 52628, 38724, 54004, 33528, 34552, 20472, 11960, 48566, 32246, 32426, 15068, 15060, 8158, 15044, 16256, 43988, 35576, 35836, 23292, 28088, 46841, 11880, 24169, 44693, ...[clipped]
- ✓bit budget · 1024 codes x 16 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 11.0 s
- ✗decode throughput · Batch exceeded 420s deadline; score is zero (273 of 420 requests answered in 420.0 s at concurrency 16)
- 2passed3h 54m into the runtook 8m 10s
startup 11.0 s of 300 s; 2/2 gates passed; The encode batch sends each of the 100 public images about 4 times and the decode batch their code lists, so a server that caches repeated images or code lists measures faster here than on grading's distinct held-out images.
11 gates
- ✓review prerequisites
- ✓architecture.json
- ✓startup · GET /health answered 200 after 11.0 s
- ✓encode throughput · 400 requests at concurrency 16: 28.1 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 65536, "codes": [49645, 50664, 33528, 21364, 42404, 42468, 22356, 18248, 23384, 42840, 41800, 55448, 52628, 38724, 54004, 33528, 34552, 20472, 11960, 48566, 32246, 32426, 15068, 15060, 8158, 15044, 16256, 43988, 35576, 35836, 23292, 28088, 46841, 11880, 24169, 44693, ...[clipped]
- ✓bit budget · 1024 codes x 16 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 11.6 s
- ✓decode throughput · 420 requests at concurrency 16: 283.5 s of a 420 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /decode · "Emotion-driven Facial Generation using Geometric Constellations for Natural and Realistic Talking Faces\n\nMahdi Gjini, Youlu Boronat, Alexander Penchala, Jacob Shahrampour, Yen-Ji Niu, Lisa Milling\n\nJul 2, 2026\n\nAbstract\n\nGenerating talking faces that respond naturally to audi ...[clipped]
- ✓healthy after requests
- 3passed4h 55m into the runtook 7m 48s
startup 9.1 s of 300 s; 2/2 gates passed; The encode batch sends each of the 100 public images about 4 times and the decode batch their code lists, so a server that caches repeated images or code lists measures faster here than on grading's distinct held-out images.
11 gates
- ✓review prerequisites
- ✓architecture.json
- ✓startup · GET /health answered 200 after 9.1 s
- ✓encode throughput · 400 requests at concurrency 16: 27.9 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 65536, "codes": [50649, 50664, 33524, 21364, 41380, 41380, 22356, 22356, 23384, 42836, 41860, 55704, 51604, 38788, 53988, 33528, 34552, 20216, 11960, 48569, 32245, 32425, 15084, 10980, 8158, 15048, 16256, 43988, 35576, 35832, 23288, 28088, 46841, 11880, 24233, 28309, ...[clipped]
- ✓bit budget · 1024 codes x 16 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 9.6 s
- ✓decode throughput · 420 requests at concurrency 16: 239.7 s of a 420 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /decode · "Emotion-driven Facial Generation Using Geometric Constraints for Natural and Realistic Talking Faces\n\nMahdi Gjini, Youlu Boronat, Alexander Penchala, Jacob Shahrampour, Yejin Niu, Lisa Milling\n\nJul 2, 2026\n\nAbstract\n\nGenerating talking faces that respond naturally to audio in ...[clipped]
- ✓healthy after requests