OCR latent compression
- Score
- 37.2
- Wallet spend
- $8.91 of $10
- Model cost
- $14.48
- Agent time
- 5h 30m
- Steps
- 414
- Tokens in / out
- 39.8M / 275k
- Self-checks
- 2 passed of 3 used
- Rank at $10 wallet
- 5th of 19
Score
37.2raw metric 0.3324
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
- 256
- diagnostic_word_f1
- 0.4438
- diagnostic_char_similarity
- 0.5103
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 codesI traced POST /encode from server.py to BatchProcessor, then to runtime.VisualTextModel.encode_batch. The image is base6
I traced POST /encode from server.py to BatchProcessor, then to runtime.VisualTextModel.encode_batch. The image is base64-decoded, converted to RGB and padded to 1024x1024 (bicubic resize, gray padding). It is then normalized to [-1,1] and passed to visual_tokenizer.visual_features, which runs only the frozen DeepSeek-OCR SAM, CLIP and projector to produce [B,256,1280] features. ResidualVisualQuantizer.encode then quantizes those features with 8 greedy residual-VQ stages of 256 entries each (argmin over learned codebooks). That gives 2048 integer codes in [0,256), i.e. 2048x8 = 16,384 bits. Serialization is a plain flatten. No OCR, language-model generation, tokenizer call or text transcription runs anywhere in /encode.
✓decoder generates text from codes onlyserver.py checks that the request has exactly 2048 ints in [0,256) and codebook_size 256. runtime.decode_batch then dese
server.py checks that the request has exactly 2048 ints in [0,256) and codebook_size 256. runtime.decode_batch then deserializes the codes and rebuilds embeddings: the selected codebook vectors are summed and passed through the learned output Linear. It adds the shared learned newline and separator embeddings plus fixed BOS and prompt-suffix embeddings, then runs the DeepSeek-V2 MoE language model with greedy argmax decoding and a KV cache. A deterministic n-gram suppression step uses only the tokens generated so far. Padding rows in the canonical batch of 16 copy the request's own codes and are thrown away. The dense-route MoE in fast_moe.py computes every row independently. No file, cache, database, socket or environment variable carries per-image information between requests. The env vars that are read are fixed config values, set at startup. There is no transcript store and no retrieval over a text corpus. Decoding is greedy, and torch is seeded at startup.
✓served weights are the trained modelserve.sh runs server.py, which loads the base checkpoint (model-00001-of-000001.safetensors) offline with AutoModel. It
serve.sh runs server.py, which loads the base checkpoint (model-00001-of-000001.safetensors) offline with AutoModel. It then injects LoRA, loads and merges adapters.safetensors, and loads quantizer.safetensors strictly into ResidualVisualQuantizer. The inventory shows codebooks [8,256,1280] plus output.weight and output.bias, which matches the residual class and codebook_size 256. The adapter inventory (lora_A/lora_B, norms, image_newline, etc.) matches lora.inject_adapters and train.py adapter_state. The trajectory shows the residual1 training run finishing at update 19968. Its best/last checkpoint hashes (adapters b28d45d1..., quantizer 05a64692...) are identical (trajectory#283). Those files were backed up to Spin storage, copied into /workspace/output and re-hashed to the same values (trajectory#284, trajectory#295), and the same hashes are recorded in evidence/checkpoint_sha256.json. The base model is declared as deepseek-ai/DeepSeek-OCR at a pinned revision, and its sha256 matches the download hash in evidence/base_download_sha256.txt and the download command in evidence/download_commands.sh. An earlier product-quantizer checkpoint briefly sat in the output directory, but it was replaced. I found no decoy or alternate model.
✓harness credentials used only by the harnessI searched the trajectory for harness credentials (ANTHROPIC, OPENAI, OPENROUTER, API_KEY, BASE_URL) and for extra harne
I searched the trajectory for harness credentials (ANTHROPIC, OPENAI, OPENROUTER, API_KEY, BASE_URL) and for extra harness CLI launches (claude -p, codex). The only matches were Spin documentation text and the participant's own scans of the output directory for leaked credentials (trajectory#204-219 range, trajectory#396-404 range). Remote compute and storage went through the Spin CLI, which is allowed. No step calls a model API with the harness credentials, copies them anywhere, or starts additional harness sessions.
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.
- 1passed1h 49m into the runtook 10m 35s
startup 8.7 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 8.7 s
- ✓encode throughput · 400 requests at concurrency 16: 18.4 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 256, "codes": [76, 161, 160, 241, 172, 147, 0, 255, 183, 25, 27, 179, 57, 255, 0, 15, 239, 207, 115, 179, 83, 255, 147, 15, 249, 250, 4, 150, 7, 21, 23, 50, 69, 192, 183, 120, 187, 43, 236, 168, 69, 192, 183, 120, 187, 9, 236, 168, 45, 254, 242, 226, 72, 64, 218, 147 ...[clipped]
- ✓bit budget · 2048 codes x 8 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 10.6 s
- ✓decode throughput · 420 requests at concurrency 16: 145.0 s of a 420 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /decode · "Attention for Control in optimization for compression for dynamics and privacy in planning for graphs\nJaan Adjei, Arya Boronat, Lawrence Matsushima, Avery Shahrampour, Ziwei Jung, Sinan Blum\nDec 19, 2026\nAbstract\nControl learning tasks that must be learned internally, a data-driv ...[clipped]
- ✓healthy after requests
- 2failed4h 32m into the runtook 10m 16s
startup 8.3 s of 300 s; 1/2 gates passed; failed decode throughput: Batch exceeded 420s deadline; score is zero (305 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 t
9 gates
- ✓review prerequisites
- ✓architecture.json
- ✓startup · GET /health answered 200 after 8.3 s
- ✓encode throughput · 400 requests at concurrency 16: 21.5 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 256, "codes": [76, 37, 133, 20, 57, 83, 189, 221, 74, 110, 14, 224, 66, 64, 122, 130, 74, 187, 126, 193, 102, 128, 209, 60, 202, 228, 78, 18, 106, 77, 189, 250, 235, 222, 10, 228, 75, 83, 12, 143, 235, 222, 10, 228, 75, 83, 12, 143, 60, 194, 202, 143, 198, 199, 187, ...[clipped]
- ✓bit budget · 2048 codes x 8 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 9.5 s
- ✗decode throughput · Batch exceeded 420s deadline; score is zero (305 of 420 requests answered in 420.0 s at concurrency 16)
- 3passed5h 23m into the runtook 7m 18s
startup 9.9 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 9.9 s
- ✓encode throughput · 400 requests at concurrency 16: 26.2 s of a 400 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /encode · {"codebook_size": 256, "codes": [76, 37, 133, 20, 57, 83, 189, 221, 74, 110, 14, 224, 66, 64, 122, 130, 74, 187, 126, 193, 102, 128, 209, 60, 202, 228, 78, 18, 106, 77, 189, 250, 235, 222, 10, 228, 75, 83, 12, 143, 235, 222, 10, 228, 75, 83, 12, 143, 60, 194, 202, 143, 198, 199, 187, ...[clipped]
- ✓bit budget · 2048 codes x 8 bits for one public image; every image must fit
- ✓healthy after requests
- ✓startup · GET /health answered 200 after 9.8 s
- ✓decode throughput · 420 requests at concurrency 16: 253.2 s of a 420 s deadline, timed from a fresh instance's readiness as in grading
- ✓POST /decode · "Electron-driven Facial Generation Using Geometric Constraints for Neural and Realistic Talking Faces\nMahdi Gjini, Youlu Boronat, Alexander Penchala, Jacob Shahrampour, Yen-Ju Niu, Lisa Milling\nJul 2, 2026\nAbstract\nGenerating talking faces that respond naturally to audio input req ...[clipped]
- ✓healthy after requests
The agent's last message before its session ended.
The H100 selfcheck passed on the frozen artifact:
- Startup: 9.9 s and 9.8 s, versus 300 s allowed.
- Encode: 15.25 images/s.
- Decode: 1.6588 images/s, versus 1.0 required.
- All 420 decode requests completed in 253.2 s.
- Bit budget, smoke checks, health, and static prerequisites passed.
I’m leaving the checked artifact unchanged and proceeding to the required two-step submission.